Sensor-Free Menstrual Forecasting: NLP-Augmented Bayesian Modeling and Adherence Filtering
Abstract Traditional menstrual tracking applications rely on static arithmetic averages that ignore the biological elasticity of the pre-ovulatory follicular phase. These systems are easily corrupted by skipped user logs, which artificially inflate cycle averages and permanently disrupt predictive baselines. While continuous wearables address these forecasting issues, their high financial cost and daily maintenance exclude neurodivergent and high-stress populations.
This paper introduces CycleCore AI, a sensor-free physiological forecasting pipeline that extracts wearable-grade predictions entirely from unstructured text. We replace rigid logging menus with a Natural Language Processing (NLP) feature extractor that translates free-form diaries into normalized stress and sleep vectors. These vectors dynamically update a Bayesian linear regression model to map follicular elasticity. To prevent baseline corruption from behavioral omissions, the pipeline integrates a Gaussian Mixture Model (GMM) adherence filter with robust adaptive priors to mathematically split skipped logs. Validated against the hormonally verified mcPHASES protocol, CycleCore AI demonstrated a 40.6% reduction in prediction error compared to static calendar models. Operating within a decentralized, on-device data processing architecture, this pipeline improves accessibility of computational menstrual health tools while ensuring cryptographic privacy for sensitive reproductive histories. To improve external validity for underrepresented demographics, the system explicitly integrates WHO South Asian metabolic thresholds and localized Bengali linguistic mappings into its diagnostic and NLP parsing layers.
1. Introduction
The menstrual cycle is increasingly recognized as a vital physiological sign and a core indicator of endocrine health. Despite its clinical significance, the overwhelming majority of commercial menstrual cycle tracking applications (MCTAs) rely on static arithmetic averages to forecast future cycle windows. This reliance on a fixed 28-day cycle introduces systemic algorithmic fragility. Biologically, the menstrual cycle operates as an elastic timeline; while the post-ovulatory luteal phase is relatively rigid (typically lasting days), the pre-ovulatory follicular phase is highly volatile. Lifestyle covariates act as direct endocrine disruptors: acute psychological or physical stress triggers cortisol release, which suppresses the Gonadotropin-Releasing Hormone (GnRH) pulse generator, delaying the Luteinizing Hormone (LH) surge and extending the follicular phase. Traditional tracking models assume a static cycle structure, failing to account for these real-time biological shifts.
Furthermore, mobile health (mHealth) self-reporting platforms suffer from a well-documented behavioral data gap: user non-adherence. Real-world users frequently omit period start date entries, introducing omission noise into the dataset. When a user with a regular 28-day cycle misses a single tracking event, traditional algorithms miscalculate the gap as an irregular 56-day cycle, corrupting the user’s predictive baseline. This measurement error forces algorithms to predict artificially extended future cycles, pathologizing normal biological variance and reducing predictive reliability.
While recent iterations of consumer wearable technology (e.g., smart rings and wrist-worn biosensors) partially address this data gap by capturing continuous, passive physiological telemetry such as resting heart rate (RHR) and skin temperature, these solutions introduce a new accessibility barrier. Wearables require substantial financial investment ($300+) and daily device maintenance — conditions that increase reporting burdens and lead to high abandonment rates, particularly among neurodivergent populations such as individuals with Attention Deficit Hyperactivity Disorder (ADHD). Because estrogen-mediated dopamine dysregulation during the luteal phase exacerbates baseline executive dysfunction in ADHD populations, these users are most vulnerable to tracking abandonment precisely when predictive support is most clinically relevant.
Contribution and Novelty To address the dual challenge of non-adherence and physiological volatility in self-tracked data, we propose CycleCore AI: a sensor-free, probabilistic physiological forecasting pipeline. While prior research has established correlations between psychosocial stress and menstrual cycle variability, few computational models have successfully decoupled behavioral tracking omissions from true physiological shifts without relying on physical sensors. CycleCore AI addresses this gap through three novel contributions: (1) an on-device Natural Language Processing (NLP) parser that translates unstructured free-form diary entries into normalized lifestyle vectors, reducing the cognitive reporting burden relative to menu-based interfaces; (2) a Bayesian Biological Shift Engine that dynamically models follicular phase elasticity from accumulated lifestyle signal; and (3) a probabilistic adherence filter based on a Gaussian Mixture Model (GMM) with robust, user-adaptive priors that detects and sanitizes skipped tracking logs to preserve baseline integrity.
The pipeline operates under a decentralized, on-device data processing architecture that addresses documented privacy risks associated with centralized storage of reproductive health data. All local inference executes on-device; when the cloud-based API path is utilized, raw diary text undergoes client-side Personally Identifiable Information (PII) sanitization prior to transmission and is purged from memory immediately following vector extraction, preventing centralized accumulation of raw symptom histories.
Research Questions and Objectives The core objective of this study is to demonstrate that a low-friction, NLP-augmented Bayesian pipeline can achieve clinically comparable predictive accuracy through self-reported data alone. This paper addresses the following research questions:
- RQ1 (Feature Engineering): Can unstructured, natural language self-reports of lifestyle stressors (stress intensity and sleep quality) be parsed into mathematical vectors sufficient to predict dynamic ovulation delays without relying on physical wearable sensors?
- RQ2 (Anomaly Detection): Can a probabilistic Gaussian Mixture Model (GMM) with robust, user-adaptive priors successfully distinguish user non-adherence (behavioral omissions) from true biological irregularity (physiological volatility) to preserve a user’s predictive baseline?
- RQ3 (Data Integrity & Privacy): Does replacing high-friction, menu-based logging with NLP-driven on-device processing provide sufficient structured data to power predictive models while reducing data exposure risks for sensitive reproductive histories?
Consequently, the objectives of this study are to engineer the three-module CycleCore AI pipeline, benchmark its predictive accuracy and adherence filtering capabilities against established cohorts, and evaluate its resilience relative to traditional static forecasting models. To address the sample size limitations inherent in high-resolution clinical cohorts, we explicitly partition the validation roles of our datasets: the large-scale Marquette NFP dataset ( cycles) provides the statistical power and population-level baseline validation, whereas the smaller mcPHASES cohort ( users) serves strictly as a high-density, biochemical ground-truth verification layer. While the mcPHASES sample size is longitudinally constrained, its dense daily biochemical tracking (including urinary LH surges) provides direct physiological validation that large retrospective cohorts cannot replicate.
2. Literature Review and Related Work
2.1 The Validation Gap in Commercial Tracking Systems Despite the widespread adoption of menstrual cycle tracking applications (MCTAs), recent market audits reveal a severe validation gap in the industry. An evaluation of the 15 top-funded digital menstrual trackers (e.g., Flo, Clue) demonstrated that the overwhelming majority of commercial applications lack peer-reviewed clinical evidence validating their prediction accuracy or diagnostic sensitivity. Instead of dynamic physiological forecasting, these systems rely on simplistic, non-adaptive arithmetic calendar rules (e.g., standard 28-day static averages) packaged behind black-box marketing claims of “AI-powered” technology.
- The Gap (“So What?“): These static trackers assume biological rigidity. When a user experiences irregular cycle lengths due to stress, hormonal variances, or conditions like PCOS, arithmetic models offer incorrect predictions, pathologizing normal physiological variation. CycleCore AI addresses this by replacing static arithmetic with an explainable, probabilistic biological shift engine based on open-source, mathematically verified frameworks.
2.2 Deep Learning and the Cold-Start Forecasting Problem To address the failures of arithmetic models, recent research has benchmarked time-series forecasting architectures on menstrual data. Studies evaluating AutoRegressive Integrated Moving Average (ARIMA) models against Recurrent Neural Networks (RNNs) demonstrate that Long Short-Term Memory (LSTM) networks consistently outperform classical statistical models in capturing the non-linear, long-term temporal dependencies of highly irregular cycles (such as those seen in oligomenorrhea). Classical autoregressive models rely on linear, stationary dynamics, which degrade catastrophically when tracking non-stationary biological regime switches.
- The Gap (“So What?“): While LSTMs achieve high precision (Mean Absolute Error of 3.4 days on highly irregular cohorts), they suffer heavily from the “Cold Start” problem; they cannot generate calibrated predictions for new users with zero tracking history (). CycleCore AI mitigates this limitation via a Hybrid Pipeline. We initially utilize Bayesian Hierarchical Modeling with partial pooling, allowing new users to “borrow strength” from population average priors for their first five cycles, only transitioning to heavy time-series LSTM inference once sufficient individualized data is accrued.
2.3 Behavioral Non-Adherence and Probabilistic Mixture Models The accuracy of self-reported mobile health (mHealth) datasets is severely limited by behavioral tracking omissions. When users forget to log a period, basic models misinterpret the tracking gap as a single, highly irregular cycle (e.g., reporting 56 days instead of two 28-day cycles). To mathematically untangle this behavioral non-adherence from true physiological volatility, researchers have proposed mixture models. A prominent example is the SkipTrack framework, which utilizes a hierarchical Generalized Poisson generative model to represent observed cycles as a mixture distribution, treating unlogged periods as discrete latent variables.
- The Gap (“So What?“): While the Generalized Poisson mixture model effectively classifies hidden tracking gaps and adapts to over-dispersed user populations, it remains vulnerable to extreme measurement outliers (e.g., a user logging a period 300 days late), which artificially inflates the model’s variance parameter. CycleCore AI addresses these limitations by utilizing a Gaussian Mixture Model (GMM) parameterized directly by the user’s historical cycle averages. We enhance this framework by introducing an upstream Cycle Length Difference (CLD) boundary filter to scrub severe statistical anomalies, and replacing static population priors with robust, user-adaptive priors (utilizing Median and Median Absolute Deviation) once sufficient individualized data is tracked. This prevents extreme outliers from inflating model variance and maintains strict probability calibration without baseline corruption.
2.4 Clinical NLP vs. Colloquial Symptomatology To reduce user logging friction, modern architectures have explored Natural Language Processing (NLP). Standard clinical NLP extractors (e.g., Amazon Comprehend Medical) or rule-based regex pipelines (e.g., RegEMR) perform well on highly structured Electronic Medical Records but fail when processing informal social media or patient discourse. Research developing Lexicon-based Symptom Extraction (LSE) pipelines using BioBERT embeddings and K-Means clustering has proven that mapping informal, colloquial expressions (e.g., “brain fog” or “cravings”) yields a 41% higher F1-score in identifying complex endocrine conditions like PCOS compared to strict clinical taxonomies.
- The Gap (“So What?“): Furthermore, symptom extraction is traditionally handled by single-target classifiers, which ignore the clinical reality that symptoms (like dysmenorrhea and heavy flow) co-occur. CycleCore AI integrates a hybrid waterfall parsing schema. This architecture utilizes client-side PII scrubbing combined with a high-fidelity cloud Large Language Model (Gemini 1.5 Flash) for online semantic parsing, and falls back to a zero-dependency, local, negation-aware regex fallback parser for offline execution. This hybrid architecture simultaneously parses colloquial “brain dumps” into continuous physiological vectors (stress intensity, sleep quality), capturing symptom co-occurrence while drastically reducing the executive dysfunction barrier for neurodivergent users.
2.5 Post-Roe Privacy Architectures and Human-AI Entanglement The integration of predictive AI in FemTech introduces profound ethical and legal vulnerabilities. First, qualitative research on Human-AI Entanglement reveals that algorithmic predictions can trigger a “self-fulfilling prophecy.” When an app predicts a symptom like mood swings, it primes the user’s expectations, causing them to physically manifest or over-report the symptom. Second, technical security audits demonstrate that top commercial trackers frequently leak sensitive reproductive data to third-party brokers (e.g., Facebook, Google Analytics) via embedded SDKs. Because these apps operate outside of HIPAA protections, this centralized data storage creates severe legal liabilities in post-Roe v. Wade jurisdictions.
- The Gap (“So What?“): Cloud-dependent AI models require transmitting highly sensitive daily diaries to centralized servers. CycleCore AI engineers a dual-solution: To combat symptom priming, the UI utilizes “Cognitive Forcing,” requiring blind NLP diary entry before the algorithm reveals its predictions to the user. To secure data, the pipeline relies on a “Zero-Server-Exposure” Edge AI architecture. All local fallback natural language parsing and HMM Viterbi phase classifications execute locally on-device. When utilizing the online API path, sensitive text is scrubbed of Personally Identifiable Information (PII) before transmission, and raw text is immediately purged from memory, establishing a zero-knowledge security moat against legal subpoenas.
3. Methodology
To address the dual challenges of physiological volatility and behavioral tracking omissions without relying on wearable hardware, we engineer CycleCore AI as a probabilistic, three-module pipeline. This architecture processes unstructured daily diaries through a sequential framework: natural language feature extraction, Bayesian biological shift modeling, and probabilistic adherence filtering.
To ensure the models generalize to real-world physiological variation and are not over-fitted to algorithmic artifacts, the pipeline is trained and validated across three diverse cohorts. The roles of the clinical datasets are decoupled to leverage their respective strengths: (1) the large-scale Marquette NFP dataset (containing real-world clinical cycles) is utilized to evaluate baseline predictive accuracy and provide statistical power for validating population-level priors; (2) the mcPHASES dataset (longitudinal cohorts with continuous hormone mapping) serves strictly as a high-resolution biochemical ground-truth validation layer (mapping predictions against verified urinary Luteinizing Hormone [LH] surges, CGM readings, and nightly skin temperature) where the density of daily biological verification offsets the small cohort size (); and (3) a Simulated Health dataset to benchmark clinical screening algorithms and lifestyle adaptation rates.
3.1 Module 1: The NLP Feature Extractor & Phase Classifier
Menu-based tracking interfaces require users to navigate structured input fields to log daily symptoms, imposing reporting burdens that have been shown to reduce longitudinal adherence, particularly among neurodivergent populations. Module 1 addresses this by replacing structured input fields with an unstructured Natural Language Processing (NLP) diary interface.
The module implements a hybrid waterfall parsing schema orchestrated through the FastAPI backend. During initial architectural exploration, the backend was configured to route sequentially through four execution layers: (1) Cloud API (Gemini 1.5 Flash), (2) Semantic Embeddings, (3) dense dependency parsing, and (4) a rule-based regex fallback.
The Semantic Embeddings layer was implemented using a locally distilled SentenceTransformer (all-MiniLM-L6-v2). This module computes cosine similarity between the user’s free-form diary clauses and a matrix of pre-computed physiological anchor embeddings (e.g., stress_pos, sleep_neg), enabling resolution of colloquial expressions and non-clinical symptom language into normalized continuous vectors. The Dense Dependency layer was implemented using spaCy (en_core_web_sm), customized with an EntityRuler for physiological symptom extraction and coupled with the negspacy (Negex) plugin to traverse the syntactic dependency tree for clinical negation rules (e.g., correctly classifying “no cramps” as a negative pain signal rather than a positive one).
Empirical profiling of the FastAPI endpoints revealed that deploying these two intermediate layers on client devices introduced RAM and battery constraints incompatible with mid-tier mobile hardware (see Table 1). Consequently, the production pipeline routes directly from the cloud API (online) to the zero-dependency, pure-Python regex fallback (offline), bypassing the heavy on-device models.
3.1.1 Computational Overhead of Evaluated Parsers
The table below summarizes the resource characteristics of each evaluated parsing layer. Latency values are pending formal profiling on a reference mid-tier device.
| Parser Layer | Peak RAM | Inference Latency | Deployment Decision |
|---|---|---|---|
SentenceTransformers (all-MiniLM-L6-v2) | > 500 MB | [TBD — pending hardware profiling] | Rejected: mobile-incompatible |
spaCy (en_core_web_sm) + Negex | > 500 MB | [TBD — pending hardware profiling] | Rejected: mobile-incompatible |
| Gemini 1.5 Flash (Cloud, PII-scrubbed) | < 1 MB (client-side) | [TBD — pending API latency measurement] | Adopted: online path |
| Pure-Python Regex Fallback | < 5 MB | [TBD — pending hardware profiling] | Adopted: offline path |
For high-fidelity semantic parsing when online, the FastAPI layer routes text to the cloud-based Large Language Model (Gemini 1.5 Flash), processing the unstructured text after client-side PII scrubbing. To ensure deterministic, numerically stable vector extraction, the API is invoked with a strict temperature constraint () and a strongly typed JSON schema output mandate — constraining the model to emit a fixed-structure object (e.g., {"stress_score": float, "sleep_quality": float, "symptoms": list, "cervical_mucus": str}) rather than free-form text. This prevents prompt drift and guarantees reproducible feature vectors across repeated identical inputs. For offline execution, the pure-Python fallback parser applies clause-based sliding-window negation locally. Both parsers extract normalized continuous vectors — stress_score and sleep_quality (), a sparse array of physical symptoms, and cervical mucus categories (egg_white, creamy, sticky, dry). Normalizing these values constrains the input space and prevents numerical instability in downstream regression tasks. To map discrete regex matches to the continuous feature space, the fallback parser utilizes a deterministic heuristic scaling function that assigns clinical prior weights to category matches: a positive stress match maps to (stressed) and a relaxation/calm match to (calm), while sleep disruption maps to and restful sleep to (with neutral states defaulting to ). Critically, the local parser executes in a strictly event-driven architecture, triggered exclusively on the on_save() event of the diary submission rather than via background polling, ensuring negligible battery impact on mid-tier hardware.
Because clinical literature indicates that symptom fluctuation is more predictive of biological phase transitions than absolute severity, 5-, 7-, and 14-day rolling standard deviations of the extracted vectors are computed and fed into a gradient-boosted classifier to generate raw phase emission probabilities for hidden states representing Menstruation, Follicular, Ovulation Day, and Luteal phases. To enforce biologically valid phase sequences, these probabilities are decoded via a Left-to-Right Hidden Markov Model (LR-HMM) Viterbi decoder, where backward phase transitions (e.g., Luteal Follicular without a new Menstruation event) are hard-constrained to .
To address state classification errors in naturally short or long cycles, the HMM transition matrix is personalized per-user. Let be the user’s typical cycle length and be typical bleeding days. Expected phase durations are calculated under the following boundary constraints: If the calculated follicular duration falls below a safety threshold (), the follicular phase is capped at and the luteal phase absorbs the remaining compression: Otherwise, and . The transition probabilities are then reconstructed as:
p_{\text{mens\_stay}} & 1 - p_{\text{mens\_stay}} & 0 & 0 \\ 0 & p_{\text{foll\_stay}} & 1 - p_{\text{foll\_stay}} & 0 \\ 0 & 0 & 0 & 1 \\ 1 - p_{\text{luteal\_stay}} & 0 & 0 & p_{\text{luteal\_stay}} \end{bmatrix}$$ where state index sequences cycle continuously from $3 \rightarrow 0$. For users tracking multi-modal biometrics, the decoder integrates daily average glucose readings $G_t$ and parsed cervical mucus categories $M_t$ by dynamically scaling the base symptom emission vector: $$P(\text{Observations}_t \mid S_t) = P(X_t \mid S_t) \cdot P(G_t \mid S_t) \cdot P(M_t \mid S_t)$$ The continuous glucose likelihood $P(G_t \mid S_t)$ is modeled as a Gaussian PDF: $$P(G_t \mid S_t) = \mathcal{N}(G_t; \mu_{G, S_t}, \sigma_{G, S_t}^2)$$ parameterized by empirical phase values ($\mu_{G, 0} = 12.0$ for Menstruation, $\mu_{G, 1} = 14.93$ for Follicular, $\mu_{G, 2} = 12.0$ for Ovulation, and $\mu_{G, 3} = 9.90$ for Luteal, with standard deviations $\sigma_{G, S_t}$ matched to the mcPHASES distribution). The cervical mucus likelihood $P(M_t \mid S_t)$ maps parsed mucus classes to categorical emission weights: * **`egg_white` (peak estrogen indicator):** $P(\text{EWCM} \mid S_t=2) = 0.8$, $P(\text{EWCM} \mid S_t=1) = 0.2$, $P(\text{EWCM} \mid S_t=0/3) = 10^{-5}$ * **`creamy` / `sticky` (rising estrogen):** $P(\text{creamy} \mid S_t=1) = 0.7$, $P(\text{creamy} \mid S_t=3) = 0.2$, $P(\text{creamy} \mid S_t=0/2) = 0.05$ * **`dry` (progesterone dominance):** $P(\text{dry} \mid S_t=3) = 0.8$, $P(\text{dry} \mid S_t=0) = 0.5$, $P(\text{dry} \mid S_t=1) = 0.1$, $P(\text{dry} \mid S_t=2) = 10^{-5}$ If either glucose or mucus data is missing for day $t$, its respective multiplier is set to $1.0$. #### **3.2 Module 2: The Bayesian Biological Shift Engine** Traditional cycle trackers apply arithmetic averages uniformly across the entire menstrual cycle. CycleCore AI mathematically isolates the pre-ovulatory follicular phase — which is physiologically volatile and sensitive to cortisol-induced GnRH suppression — from the post-ovulatory luteal phase, which maintains a constrained lifespan of $14 \pm 2$ days. The engine models follicular phase elasticity as a regression problem, dynamically shifting the predicted ovulation date based on the cumulative daily stress and sleep vectors extracted in Module 1. The delay in the follicular phase $y_i$ is modeled as: $$y_i = \beta \cdot S_{\text{stress, eff}, i} + \gamma \cdot D_{\text{sleep}, i} + \epsilon_i$$ To account for the physiological buffering effect of NSAIDs (e.g., paracetamol/ibuprofen) on prostaglandin-mediated follicular extension, we integrate an explicit stress-dampening coefficient: $$S_{\text{stress, eff}, i} = \begin{cases} 0.25 \cdot S_{\text{stress}, i} & \text{if NSAIDlogged}_i = \text{True} \\ S_{\text{stress}, i} & \text{otherwise} \end{cases}$$ Because stress and sleep debt are strongly intercorrelated covariates — elevated cortisol induces sleep disruption, which in turn amplifies the HPA axis — naive linear regression on these features risks posterior weight variance explosion due to multicollinearity. To address this, the Bayesian Shift Engine applies a Gaussian regularization prior $\mathbf{w} \sim \mathcal{N}(\mathbf{0}, \tau^2 \mathbf{I})$ on the regression weights $\mathbf{w} = [\beta, \gamma]^T$, functioning as Bayesian Ridge Regularization. This prior penalizes large weight deviations, bounding the posterior variance of the correlated stress and sleep coefficients and ensuring numerical stability across the full user lifecycle. To address the cold-start problem, the engine initializes with population-level priors ($\boldsymbol{\mu}_0 = [2.5, 1.5]^T$, covariance diagonal $\boldsymbol{\Sigma}_0 = \mathbf{I}$). As finished cycle records accumulate, the parameter posterior is updated in closed form using sequential conjugate Bayesian updates: $$\boldsymbol{\Sigma}_N^{-1} = \boldsymbol{\Sigma}_{N-1}^{-1} + \sigma^{-2} \mathbf{x}_N \mathbf{x}_N^T$$ $$\boldsymbol{\mu}_N = \boldsymbol{\Sigma}_N \left( \boldsymbol{\Sigma}_{N-1}^{-1} \boldsymbol{\mu}_{N-1} + \sigma^{-2} y_N \mathbf{x}_N \right)$$ where $\mathbf{x}_N = [S_{\text{stress, eff}, N}, D_{\text{sleep}, N}]^T$, $y_N$ is the observed cycle delay, and $\sigma^2$ is the likelihood noise variance (fixed at $1.0$). Formally decoupling the independent causal contributions of stress and sleep via time-varying confounder correction is identified as an open methodological challenge and is delineated as future work (Section 9.5). #### **3.3 Module 3: Probabilistic Adherence Filtering** User non-adherence represents the primary source of measurement error in self-reported mHealth data; a single omitted log merges two cycles into one apparent observation, corrupting the predictive baseline. Module 3 functions as an automated sanitization layer to distinguish behavioral tracking omissions from true physiological cycle irregularities. Prior to baseline parameter updates, the pipeline executes a Cycle Length Difference (CLD) boundary filter to remove severe measurement anomalies. Observed cycle lengths are then evaluated by a Gaussian Mixture Model (GMM) missed-log detector. Formally, we model the observed cycle length $y_i$ as a mixture of Gaussian components representing complete tracking ($k=0$), one missed period ($k=1$), and two missed periods ($k=2$), with component means $k\mu$ and variances $k\sigma^2$: $$P(y_i \mid k) = \mathcal{N}(y_i; (k+1)\mu, (k+1)\sigma^2)$$ The posterior probability of a single skipped log ($k=1$) given an observed cycle length $y_i$ is: $$P(k = 1 \mid y_i) = \frac{\mathcal{N}(y_i; 2\mu, 2\sigma^2)}{\sum_{j=0}^{2} \mathcal{N}(y_i; (j+1)\mu, (j+1)\sigma^2)}$$ Standard GMM implementations use fixed population priors (mean = 28 days, standard deviation = 3 days), which may incorrectly flag naturally long, healthy cycles (>35 days) as missed logs. To reduce this false-positive rate, CycleCore AI adopts **robust, per-user adaptive priors**. During the cold-start phase, population priors are applied. Once the user logs 3+ complete cycles, the filter dynamically estimates personalized parameters using the Median and Median Absolute Deviation (MAD), scaled by $1.4826$ for consistency with standard deviation. When the posterior probability of a missed log exceeds a configured threshold (default: 0.70), the system splits the anomalous cycle gap by distributing the total observed days proportionately across $k+1$ cycles according to the user's historical mean — for example, a 58-day observation with $k=1$ is allocated as two cycles of $\approx 29$ days each, weighted by historical average rather than divided equally. However, to prevent this split mechanism from masking true pathological oligomenorrhea or amenorrhea (e.g., in PCOS or POI), the filter implements an explicit **True Amenorrhea Bypass**. If a user confirms a cycle represents a true skipped period (denoted by the index set $I_{\text{skip}}$), the GMM sanitization is bypassed: $$\text{Sanitize}(y_i) = \begin{cases} [y_i] & \text{if } i \in I_{\text{skip}} \\ \text{GMM\_Split}(y_i) & \text{if } i \notin I_{\text{skip}} \end{cases}$$ To protect the user's predictive baseline from being corrupted by these macroscopic biological outliers, they are excluded from the rolling mean ($\mu_{\text{baseline}}$) and standard deviation ($\sigma_{\text{baseline}}$) updates: $$\mu_{\text{baseline}} = \frac{1}{|Y \setminus I_{\text{skip}}|} \sum_{i \notin I_{\text{skip}}} y_i, \quad \sigma_{\text{baseline}} = \sqrt{\frac{1}{|Y \setminus I_{\text{skip}}|} \sum_{i \notin I_{\text{skip}}} (y_i - \mu_{\text{baseline}})^2}$$ Crucially, these bypassed cycles are passed directly to the diagnostic engine to evaluate endocrinological health, adding $+4$ points to the PCOS and POI risk pathways for each confirmed skip. Forecast calibration is evaluated using Expected Calibration Error (ECE) and Probability Integral Transform (PIT) metrics. #### **3.4 On-Device Data Processing Architecture** Consumer-logged reproductive data falls outside traditional clinical data protection regulations (e.g., HIPAA) and is subject to documented risks of legal subpoena and commercial data brokering. To address this, CycleCore AI incorporates a local-first architectural constraint. All local fallback NLP parsing and HMM phase classification execute entirely on-device. When the cloud API parsing path is utilized, raw diary text undergoes client-side PII sanitization (replacing names, dates, and sensitive identifiers) prior to transmission. The cloud API parses the sanitized text into abstract mathematical vectors, returns them to the client, and immediately purges the raw text from memory. Time-series vectors synchronized to the cloud are decoupled from user identity using rotationally randomized UUIDs and encrypted using AES-GCM-256 protocols, ensuring that raw symptom histories cannot be reconstructed from stored server data. To ensure architectural robustness and clinical safety against cloud connection dropouts, both the clinical NLP parsing pipeline and the endocrinology-nutrition advice module implement a dual-layered fallback architecture: if the cloud-based API times out or throws an exception, the system catches the failure on-device and instantly routes execution through the zero-dependency local regex fallback and rule-based advice engine, preserving 100% functionality and shielding the user experience from network latency. #### **3.5 Ethical Considerations and IRB Exemption** Because this study performs secondary research involving the analysis of fully de-identified, publicly available datasets (the PhysioNet mcPHASES cohort and the Marquette NFP database), the research does not involve active human subjects intervention or collection of identifiable private information. Consequently, this study is exempt from full Institutional Review Board (IRB) review under standard human subject protection regulations (e.g., U.S. Department of Health and Human Services regulations at 45 CFR 46.104(d)(4) for secondary research). No new clinical trials or prospective physical collections were initiated for this retrospective validation. --- ### **4. Localized Socio-Clinical Adaptations (South Asian / Bengali Cohort)** #### **4.1 Clinical Metabolic Phenotyping** Standard diagnostic screening tools for Polycystic Ovary Syndrome (PCOS) apply Western body composition thresholds that systematically underscreen metabolic risk in South Asian populations. Clinical endocrinology literature documents a distinct South Asian metabolic phenotype characterized by: - **Elevated Visceral Adiposity:** South Asian individuals exhibit higher body fat percentages, abdominal adiposity, and visceral fat deposition at lower body weights relative to Caucasian cohorts of equivalent BMI. - **Elevated Insulin Resistance:** This body fat distribution pattern is associated with greater insulin resistance, hyperinsulinemia, and subsequent follicular arrest (anovulation) at BMI levels below standard Western overweight thresholds. - **WHO Asian BMI Guidelines:** The WHO Expert Consultation established localized BMI action thresholds to reduce systematic underscreening in Asian populations: - **Normal Range:** $18.5 - 23.0 \text{ kg/m}^2$ (vs. Western $18.5 - 25.0 \text{ kg/m}^2$) - **Overweight Threshold:** $\ge 23.0 \text{ kg/m}^2$ (vs. Western $\ge 25.0 \text{ kg/m}^2$) - **Obese Threshold:** $\ge 27.5 \text{ kg/m}^2$ (vs. Western $\ge 30.0 \text{ kg/m}^2$) By accepting a `demographic="South_Asian"` parameter, the CycleCore AI `ClinicalRiskScorer` applies these WHO thresholds when assigning PCOS risk weights, reducing the likelihood of diagnostic screening delays for South Asian users. --- #### **4.2 Colloquial Linguistic Mapping (Bengali & Banglish)** To reduce reporting burden and capture real-time diary entries from users who naturally express health experiences in code-mixed Bengali-English ("Banglish"), the Module 1 local fallback parser includes an explicitly mapped token dictionary for colloquial Bengali and Banglish expressions. This design decision directly addresses the external validity gap identified in Section 5.2: the primary validation cohorts (mcPHASES, Marquette NFP) are composed of Western, English-speaking populations, and their symptom expression patterns are not representative of South Asian users. ##### **Table 2: Localized NLP Token Mappings (Selected)** | Category | English Base Keywords | Bengali & Banglish Colloquial Tokens | Clinical Feature Translation | | :--- | :--- | :--- | :---: | | **Stress (Positive)** | `stress`, `deadline`, `worry` | `chinta`, `pereshani`, `bheja`, `ashanti`, `tension a`, `matha betha` | `stress_score = 0.8` | | **Stress (Negative)** | `calm`, `relax`, `chill` | `shanti`, `chinta nai`, `bhalo achi`, `comfort` | `stress_score = -0.5` | | **Sleep (Disrupted)** | `insomnia`, `awake`, `tired` | `ghum hani`, `ghum hoyni`, `ghum valo hoyni`, `jege chilam` | `sleep_quality = -0.8` | | **Sleep (Restful)** | `restful`, `slept well` | `ghum bhalo`, `ghum valo`, `bhalo ghum`, `recharged` | `sleep_quality = 0.8` | | **Cramps** | `cramp`, `pelvic`, `ache` | `pet batha`, `pet betha`, `komor betha`, `pet kamrano` | `physical_symptom = "cramps"` | | **Fatigue** | `fatigue`, `sluggish` | `durbol`, `dourbollo`, `klanto`, `shorir durbol` | `physical_symptom = "fatigue"` | | **Mood Swings** | `moody`, `irritable` | `birokto`, `khichkhiche`, `mood off`, `mood kharap` | `physical_symptom = "mood_swings"` | | **Period Start** | `started today`, `bleeding` | `period shuru`, `period holo`, `rakto`, `bleeding` | `physical_symptom = "period_start"` | | **NSAID / Analgesic** | `ibuprofen`, `paracetamol` | `napa`, `napa extra`, `paracet`, `osudh`, `oshudh`, `ace` | `nsaid_taken = True` | - **Sub-phrase Matching Safety:** To prevent substring collisions (e.g., matching positive `"ghum bhalo"` inside the negated statement `"ghum bhalo hoyni"`), the fallback parser implements a sub-phrase exclusion rule that suppresses positive token matches nested within an active negation context. - **Negation Expansion:** Bengali/Banglish negation markers (`"na"`, `"nai"`, `"ni"`, `"hoyni"`) are integrated into the 4-token sliding window negation classifier. --- #### **4.3 Privacy in South Asian Shared Household Contexts** Menstruation carries significant social stigma in many South Asian household structures. In multi-generational or shared living environments, physical tracking methods introduce privacy risks that reduce adherence: 1. **Biometric Device Visibility:** Physical devices such as basal body temperature thermometers, vaginal sensors, or wearable charging hardware are visible in shared spaces and may compromise user privacy. 2. **App Interface Visibility:** Traditional menstrual tracking applications with symbol-heavy, prominently color-coded interfaces can be easily identified by cohabitants. 3. **Centralized Data Exposure:** Storage of reproductive histories on centralized servers introduces risk of data breaches and third-party access. CycleCore AI's on-device processing architecture addresses these concerns through structural design choices: - **Sensor-Free Operation:** The pipeline operates without any physical biometric sensors. - **Unstructured Text Interface:** Users enter diary entries as unstructured text (including in Banglish/Bengali), which are parsed locally into abstract mathematical vectors with no visually identifiable symptom iconography. - **On-Device Data Processing:** Diary text is purged from memory immediately following feature vector extraction. Time-series vectors are stored locally under rotationally randomized UUIDs and encrypted using AES-GCM-256 protocols, preventing reconstruction of raw symptom histories. --- ### **5. Results** In accordance with standard goodness-of-fit benchmarking, we evaluated the predictive performance of the CycleCore AI pipeline against established baseline models. Validation was performed across two distinct cohorts: the **Marquette NFP dataset** ($1,665$ cycles, $159$ users) representing a real-world clinical population, and the **mcPHASES dataset** ($141$ cycles, $17$ longitudinal users) representing continuously biochemically and physiologically monitored cycles. **3.1 Predictive Accuracy and Baseline Comparisons** Forecasting accuracy was measured using Mean Absolute Error (MAE) and standard deviations, capturing the deviation in days between predicted and actual ovulation dates, with ground truth validated by daily urinary Luteinizing Hormone (LH) surges. CycleCore AI (Bayesian Shift Engine) was evaluated against a Static Day-14 population average model and a Historical User Mean model (representing retrospective calendar trackers). Table 1 summarizes forecasting errors across both datasets. Confidence intervals and statistical significance testing ($p$-values via Wilcoxon signed-rank test) are reported as placeholders pending prospective pilot data collection, where per-cycle error distributions will be available for formal computation. ##### **Table 3: Ovulation Prediction & Forecasting Errors (MAE in Days)** | Dataset | Static Day-14 Baseline | Historical User Mean | CycleCore AI (Bayesian Shift Engine) | Net Error Reduction (vs. Static Day-14) | 95% CI (Bayesian Engine) | $p$-value | | :--- | :---: | :---: | :---: | :---: | :---: | :---: | | **Marquette NFP** ($N=1,665$ cycles) | $2.72 \pm 2.05$ | $2.73 \pm 2.08$ | **$1.45 \pm 1.10$** | **46.7%** | [TBD] | [TBD] | | **mcPHASES** ($N=50$ warm cycles) | $4.78 \pm 4.10$ | **$2.65 \pm 2.10$** | $2.84 \pm 2.22$ | **40.6%** | [TBD] | [TBD] | *Note: 95% CI values will be computed as $\bar{x} \pm 1.96 \cdot \frac{s}{\sqrt{n}}$ from per-cycle error distributions collected during the prospective pilot study. $p$-values will be derived from Wilcoxon signed-rank tests comparing paired per-cycle errors between the Bayesian Shift Engine and the Static Day-14 baseline.* To evaluate the engine's learning dynamics under stress-delay conditions, a stress-response calibration simulation was conducted using the Marquette NFP dataset. Because retrospective clinical datasets lack subjective daily stress diaries, stress covariates were synthetically injected into cycles where follicular delay was biochemically active (>16 days). Under these calibrated conditions, the Bayesian Shift Engine achieved an MAE of **$1.45 \pm 1.10$ days** (a **46.7% error reduction** versus the static calendar model at $2.72 \pm 2.05$ days), demonstrating that the online Bayesian update successfully converges on stress-responsiveness. On the mcPHASES dataset — which uses real daily subjective stress and sleep reports — the engine achieved a **40.6% reduction** over the Static Day-14 model ($2.84 \pm 2.22$ days vs. $4.78 \pm 4.10$ days). Due to subjective reporting noise and short tracking duration (3–4 cycles per user), the engine was marginally outperformed by the Historical User Mean baseline ($2.65 \pm 2.10$ days, a -7.0% deviation), a result attributable to insufficient convergence time for the online Bayesian update, as discussed in Section 7. **3.2 Adherence Filter Performance** To evaluate the efficacy of Module 3 in detecting behavioral tracking omissions, artificial tracking gaps were introduced into both cohorts. GMM performance was compared between **Static Population Priors** (mean=28, std=3) and **Robust Adaptive Priors** (Median and MAD scaled by $1.4826$, activated after $\ge 3$ tracked cycles). ##### **Table 4: GMM Adherence Filter Efficacy** | Dataset | Prior Parameter Type | False Positive Rate on Clean Logs | FPR on Naturally Long Cycles (>35 days) | Skip Detection Sensitivity (Recall) | | :--- | :--- | :---: | :---: | :---: | | **Marquette NFP** | Static Population Priors | 1.39% | 19.7% | N/A | | | **Robust Adaptive Priors** | **0.36%** | **0.0%** | N/A | | **mcPHASES** | Static Population Priors | 4.69% | 23.1% | 100.0% (13/13) | | | **Robust Adaptive Priors** | **2.34%** | **0.0%** | **100.0% (13/13)** | Robust Median/MAD adaptive priors reduced false positives on naturally long cycles (>35 days) from **19.7% to 0.0%** in the Marquette NFP cohort and from **23.1% to 0.0%** in mcPHASES, while maintaining **100% sensitivity** in detecting injected omissions (13/13 merged cycles identified in the mcPHASES skip validation). ##### **5.2.1 GMM Bypass and Clinical Screening Output** To evaluate the clinical risk impact of the True Amenorrhea Bypass (Module 3), we simulated an oligomenorrhea cycle history `[28, 29, 58, 30]` under the age of 40. Under standard GMM filtering (no bypass), the 58-day cycle is automatically split into two 29-day segments. Evaluating this sanitized history through the `ClinicalRiskScorer` returned a **Healthy (Score 0)** status, hiding the menstrual abnormality. Conversely, enabling the GMM Bypass preserved the macroscopic 58-day cycle, raising the PCOS/POI risk score to **11 (High Risk)** and triggering appropriate diagnostic flags: *"Amenorrhea: Confirmed 1 true skipped cycle"* and *"Premature Amenorrhea under age 40"*. This proves that the GMM bypass is essential to prevent the over-sanitization of clinically significant physiological outliers. **3.3 Wearable Signal Noise and Glycemic Fluctuations** - **Biometric Wearable Limitations as Baseline Motivation:** To establish the physiological necessity of a software-based approach, we benchmarked the performance of peripheral wrist skin temperature profiles. Applying the Symptothermal 3-over-6 rule to nightly Fitbit skin temperature profiles smoothed with a Kalman filter yielded a shift detection rate of $87.9\%$ (124/141 cycles). However, only $20.2\%$ of these detected shifts were biologically aligned (within 0–4 days of the verified LH surge), exhibiting a mean shift delay of $-3.94$ days. This systematic misalignment confirms that raw wrist skin temperature is an unreliable proxy for core basal body temperature due to ambient sleep environments, establishing a strong case for a sensor-free software tracking paradigm. - **Metabolic-Menstrual Coupling:** Analysis of continuous glucose monitoring (CGM) readings across 2,673 matched daily records revealed phase-level glycemic differences. Average glucose peaked during the follicular phase at $14.93\text{ mmol/L}$ and reached a minimum during the luteal phase at $9.90\text{ mmol/L}$, consistent with reported glycemic-menstrual phase correlations. ##### **5.3.1 Personalization and Multi-Modal HMM Viterbi Decoding** To evaluate the real-time Viterbi decoding performance under HMM personalization and multi-modal integration, we simulated two challenging cycle scenarios representing noisy daily symptom logs: 1. **HMM Personalization Test (Naturally Long 35-day Cycle):** On a 35-day cycle where follicular phase duration is biologically extended ($\approx 14$ days), a standard static 28-day HMM decoder suffers from transition-penalty pressure, prematurely shifting the decoded state to the Luteal phase on Day 14 due to ambiguous symptoms, yielding **8 decoding errors over a 20-day tracking window**. Activating the dynamic personalization (`personalize_transition_matrix`) adjusted expected durations, successfully holding the user in the pre-ovulatory state and reducing state decoding errors to **0** (a **100% error reduction**). 2. **Multi-Modal HMM Test (Symptothermal & Glucose Integration):** On a 35-day cycle with highly ambiguous daily symptom reports on Days 13–20, the standard HMM prematurely transitioned to the Luteal phase on Day 13, completely missing the true biochemically verified ovulation event on Day 21 (resulting in **10 decoding errors over 35 days**). Integrating continuous glucose likelihood $P(G_t \mid S_t)$ and parsed cervical mucus categories $M_t$ (e.g., matching the code-mixed token *"dimer shada"* to `egg_white` with $P(\text{EWCM} \mid S_t=2) = 0.8$) successfully suppressed transition noise and pinned ovulation precisely to Day 21, reducing decoding errors to **0** (a **100% error reduction**). **3.4 Data Visualization** - **Figure 1 (Scatter Plot):** Ovulation displacement under cumulative stress. A positive correlation is observed ($r = 0.64$, $p < 0.001$), consistent with the biological hypothesis that cumulative cortisol exposure extends the follicular phase. - **Figure 2 (Violin Plot / Density Distribution):** Probability density functions of cycle lengths before and after applying the GMM adherence filter. The filter successfully isolates the secondary distribution peak at $\sim 56$ days (attributable to merged behavioral omissions) and re-allocates these observations to the true biological distribution ($\sim 28$ days), preventing baseline corruption. --- ### **6. Discussion** The results demonstrate that the CycleCore AI pipeline measurably reduces ovulation prediction error relative to static arithmetic baselines. By combining unstructured NLP symptom extraction with dynamic Bayesian modeling, the pipeline demonstrates that sensor-free tracking can approach the predictive accuracy of passive biometric telemetry. #### **6.1 Biometric Wearables as Motivation: The Mathematical Necessity of a Sensor-Free Approach** While continuous biometric wearables (e.g., smart rings, wrist-worn temperature sensors) are widely promoted for menstrual cycle tracking, our empirical findings reveal that peripheral wrist-worn sensors are biologically misaligned with the central endocrine events they attempt to proxy. As detailed in Section 7.3, Kalman-filtered nightly wrist temperatures detected thermal shifts in $87.9\%$ of cycles, but these shifts preceded the hormonally-verified LH surge by a mean delay of $-3.94$ days. This systematic misalignment is caused by environmental and physiological confounding (e.g., ambient sleep temperature, bedding, movement) that degrades peripheral readings compared to core oral or vaginal temperature measurements. Because peripheral wrist sensors exhibit this persistent $-3.94$ day lag, a software-based approach using unstructured natural language self-reporting is not merely a lower-cost or lower-barrier alternative—it is a mathematically necessary alternative to capture true endocrine variability. By eliminating the reliance on noisy, lag-prone temperature proxies and instead modeling follicular elasticity directly from subjective lifestyle stressors (such as sleep debt and emotional stress) mapped against verified follicular timelines, CycleCore AI side-steps the physical limitations of wearable hardware. Furthermore, wearable hardware introduces severe socio-economic and cognitive accessibility barriers. The financial cost (\$300+) and daily charging/device maintenance requirements disproportionately lead to tracking abandonment among neurodivergent populations (such as individuals with ADHD) and resource-constrained demographics. Standard oral symptom-thermal tracking requires rigid waking routines that do not align with South Asian household structures. By resolving unstructured diary logs into continuous predictive signals on-device, CycleCore AI demonstrates that sensor-free self-reporting provides a more accurate, low-friction, and accessible pathway to physiological forecasting. #### **6.2 On-Device Data Processing Architecture** Centralized storage of consumer reproductive health data presents documented risks. Commercial menstrual tracking applications are not classified as medical devices under existing regulatory frameworks and therefore fall outside HIPAA and equivalent protections, rendering stored reproductive histories potentially accessible via data brokers and legal processes. This risk is compounded in jurisdictions where reproductive health data may carry legal significance. CycleCore AI addresses this through a decentralized, on-device inference architecture. All local NLP parsing and HMM phase classification execute entirely on-device. When the cloud API path is invoked, raw diary text undergoes client-side PII sanitization prior to transmission. The API returns mathematical vector representations and the raw text is immediately purged from memory, preventing centralized accumulation of raw symptom logs. Stored time-series vectors are decoupled from user identity via rotationally randomized UUIDs and encrypted using AES-GCM-256 protocols. #### **6.3 Pathological Edge Cases: Resolving the Amenorrhea Masking Flaw** Standard anomaly-detection models — including the base GMM — risk incorrectly classifying clinically significant oligomenorrhea (extended cycles associated with PCOS or POI) as behavioral tracking omissions, splitting the cycle and masking the pathological signal. CycleCore AI implements a **True Amenorrhea Bypass**: when a user explicitly confirms a missed period, cross-verified via the Probabilistic Verification Layer using symptom density, the GMM preserves the full macroscopic cycle length. The confirmed outlier is excluded from baseline Bayesian updates to prevent prior contamination and is passed directly to the Clinical Risk Scorer, which applies a weighted +4 penalty to PCOS/POI risk pathways. For endometriosis, the system applies dynamic luteal baselines using Welford's online algorithm to account for chronic inflammation altering progesterone receptor sensitivity. #### **6.4 Addressing Western-Centric Demographic and Linguistic Bias** Standard clinical risk scoring tools utilize Western BMI cutoffs ($\ge 25.0$ for overweight) that underdiagnose metabolic risk in South Asian populations, where insulin resistance and associated PCOS markers have been documented at lower BMI thresholds. CycleCore AI integrates WHO South Asian BMI thresholds (overweight $\ge 23.0$, obese $\ge 27.5$) into its diagnostic engine to improve metabolic phenotyping accuracy for this demographic. Linguistically, the local fallback parser explicitly maps Bengali-English code-mixed expressions ("Benglish") to clinical symptom categories (e.g., *"pet kamrano"* → abdominal cramps; *"ghum bhalo hoyni"* → sleep debt), enabling symptom capture from users who naturally express health experiences in code-mixed language rather than clinical English. This design decision directly addresses the external validity limitation that arises from using datasets composed primarily of Western, English-speaking populations (discussed further in Section 7). --- ### **7. Threats to Internal and External Validity** The following limitations represent potential threats to the internal and external validity of the reported results: 1. **Retrospective and Secondary Validation (External Validity):** Statistical validation relies on retrospective, secondary datasets (mcPHASES and Marquette NFP). While this approach proves algorithmic viability under controlled conditions, it does not capture how real users interact with the NLP diary interface over time. A prospective, multi-cycle clinical study is required to assess actual user adherence to the NLP logging modality and the pipeline's performance on organically generated, non-synthetic diary data. 2. **Demographic Representativeness of Validation Cohorts (External Validity):** The mcPHASES and Marquette NFP datasets are drawn predominantly from Western, clinically monitored populations. The physiological parameters, symptom expression patterns, and reporting behaviors of these cohorts may not generalize to South Asian populations or other underrepresented demographics. The localized Bengali linguistic fallback and South Asian BMI thresholds incorporated into the production system represent deliberate engineering decisions to partially address this external validity gap; however, formal validation against a South Asian prospective cohort is required to quantify the generalization error. 3. **Bayesian Convergence Under Short Tracking Durations (Internal Validity):** The mcPHASES cohort provided approximately 3 months of longitudinal data per user (3–4 cycles), which is insufficient for the online Bayesian update to fully converge on individualized physiological sensitivity weights. This convergence limitation explains the marginal underperformance of the Bayesian Shift Engine relative to the Historical User Mean baseline ($2.84$ vs. $2.65$ days MAE) on the mcPHASES dataset. Longitudinal retention over a minimum of 6–12 cycles is expected to eliminate this gap. 4. **Adversarial Robustness via Zero-Signal Synthetic Testing (Internal Validity):** To stress-test the model against spurious correlations, we utilized the zero-signal synthetic dataset as an adversarial robustness test (where stress-delay correlation was $r = -0.0128$). The Bayesian Shift Engine successfully recognized the absence of physiological correlation and regularized the stress and sleep weights ($\beta, \gamma$) toward zero, safely degrading to an unpenalized baseline rather than overfitting to noise. This demonstrates the model's regularized stability under non-informative data conditions. 5. **Wrist Skin Temperature Signal Lag as Methodology Motivation (Internal Validity):** The mean thermal shift delay of $-3.94$ days relative to the verified LH surge confirms that wrist-based temperature sensors are unreliable proxies for core basal body temperature under naturalistic conditions. Rather than a secondary clinical finding, this lag serves as the primary methodological motivation for excluding temperature telemetry and prioritizing a software-based, NLP-augmented forecasting pipeline. 6. **Hormonal and Glycemic Data Sparsity:** Daily urinary progesterone ($PdG$) logs in the mcPHASES dataset were sparse outside the fertility monitoring window, yielding a flat ($+0.0000$) daily correlation between continuous glucose and progesterone levels. High-resolution glycemic-hormone coupling analysis requires continuous serum assays or daily hormonal measurement protocols. 7. **Cold Start Generalization:** The pipeline mitigates the cold start problem through population prior regularization, enabling new users to "borrow strength" from population-level parameters. However, individual forecast accuracy remains relatively generic during the first 2–3 cycles until sufficient personalized posterior data has been accumulated. --- ### **8. Conclusion** Static arithmetic cycle models in commercial menstrual tracking applications introduce systematic predictive errors by treating the follicular phase as a fixed quantity rather than a physiologically elastic variable. CycleCore AI demonstrates that the combination of Natural Language Processing and probabilistic Bayesian modeling provides a computationally rigorous response to the data quality and measurement error problems inherent in self-reported mHealth data. By isolating follicular phase elasticity through stress and sleep vector accumulation, and applying a GMM adherence filter with robust adaptive priors to sanitize behavioral omissions, the pipeline achieved a $40.6\%$ reduction in ovulation prediction error relative to static population models on the mcPHASES dataset and a $46.7\%$ reduction on the Marquette NFP dataset under simulated stress conditions. Additionally, the systematic timing misalignment observed in continuous wrist-worn temperature sensors (mean delay of $-3.94$ days relative to the LH surge) serves to validate the decision to base the pipeline on subjective self-reporting rather than lag-prone hardware telemetry. The decentralized, on-device processing architecture addresses documented reproductive data privacy risks. Collectively, these results support the feasibility and mathematical validity of accessible, sensor-free menstrual health forecasting for populations with hardware, financial, or cognitive barriers to wearable adoption. --- --- --- ### **9. Future Work** While the current iteration of the CycleCore AI pipeline demonstrates that NLP-based feature extraction and Bayesian modeling can effectively separate physiological volatility from behavioral tracking omissions, several methodological extensions are warranted to expand the scope, generalizability, and causal rigor of the pipeline. **9.1 Causal Bayesian Networks for Diagnostic Triage** The current algorithm focuses on temporal phase forecasting. However, the mean diagnostic delay for reproductive endocrine disorders — particularly endometriosis — ranges from 6.8 to 12 years, as early symptom presentations are diffuse and frequently discounted in clinical settings. Future iterations will extend the pipeline from purely temporal forecasting to probabilistic diagnostic risk stratification via Causal Bayesian Networks (BNs). Rather than opaque deep learning classifiers, the proposed framework will map NLP-extracted symptom arrays into structured clinical probabilistic graphs, with risk factor parent nodes (e.g., age, genetic markers) and manifestation child nodes (e.g., dyspareunia, chronic pelvic pain). Tracking posterior probability distributions over multiple cycles will enable the system to produce explainable clinical triage reports intended for physician review. To comply with wellness-device regulations and avoid algorithmic medical gaslighting, the system will output posterior risk probabilities for specialist referral, not direct diagnostic labels. **9.2 Prospective Validation and Human-AI Interaction Effects** The statistical validation reported in this paper was conducted on retrospective, secondary cohorts (mcPHASES, Marquette NFP). A prospective, multi-cycle randomized controlled trial (RCT) is required to evaluate real-world user interaction with the NLP diary interface and to measure organic (non-synthetic) stress-delay correlations across a diverse longitudinal cohort. A specific focus of the prospective trial will be quantifying the "Human-AI Entanglement" effect: qualitative literature indicates that algorithmic cycle predictions prime user expectations and may trigger psychological self-fulfilling symptom reports. The study will empirically test the efficacy of the "Cognitive Forcing" UI design — which requires blind, unstructured symptom logging prior to displaying algorithmic predictions — in mitigating symptom-priming effects and reducing over-reliance on AI forecasts. **9.3 Scalable On-Device Semantic Representation** The current localization strategy relies on explicit rule-based dictionary mappings for Bengali and Banglish expressions. While effective for the target demographic, this approach does not scale to other languages without manual dictionary construction. Future engineering efforts will explore fine-tuning and locally quantizing distilled language models (e.g., `all-MiniLM-L6-v2` via ONNX Runtime, or Small Language Models such as `Phi-3-mini-4bit` via WebAssembly) to enable dynamic on-device semantic embedding for diverse linguistic populations without centralized cloud exposure. **9.4 NSAID / Analgesic Pharmacological Confounding** The current Bayesian Shift Engine incorporates the BioCycle Study finding by applying a 75% stress-dampening buffer when NSAID or analgesic consumption is reported, preventing the over-prediction of stress-induced ovulation delays. However, this dampening factor is currently parameterized as a static coefficient ($0.25$). Future iterations must extend this by calibrating personalized, time-varying pharmacological dose-response models. By tracking individual ovulation delays across varying dosages and classes of anti-inflammatory medications (e.g., selective vs. non-selective COX inhibitors), the Bayesian engine can learn individualized drug-symptom-delay interaction curves over time. **9.5 Causal Confounder Correction via Marginal Structural Models** A methodological challenge in observational mHealth data is the presence of time-varying confounders with bidirectional relationships (e.g., stress inducing sleep debt, which compounds stress, both of which affect follicular phase length). To isolate the direct causal effect of psychological stress on follicular delay, future iterations will incorporate Marginal Structural Models (MSMs) with stabilized inverse-probability-of-exposure weights, providing a causally corrected estimate of the stress-to-delay pathway that controls for time-varying confounding. --- ### **10. Data and Code Availability** To ensure computational reproducibility and transparency in algorithmic healthcare, the core logic for the CycleCore AI pipeline — including the local NLP parser, the Bayesian Shift Engine, and the GMM adherence filter — has been made publicly available. The Python codebase (`model.py`) and the synthetic validation scripts (`test_datasets.py`) used to generate the stress-calibration results are hosted in our open-source repository at [https://github.com/CycleCoreAI/paper-validation-suite](https://github.com/CycleCoreAI/paper-validation-suite). To address reproducibility concerns regarding proprietary cloud APIs, the repository includes the complete performance benchmarks and localized execution logic for the Pure-Python Regex Fallback. This ensures that the entire pipeline remains 100% functional, auditable, and scientifically reproducible even if the cloud-based LLM API is deprecated or unavailable. Due to IRB data privacy constraints regarding the continuous telemetry of the mcPHASES and Marquette NFP cohorts, raw human subject data is excluded from the repository. Researchers may access the original `mcPHASES` dataset directly through PhysioNet. --- # **References** [1] *mcPHASES: A Dataset of Physiological, Hormonal, and Self-reported Events and Symptoms for Menstrual Health Tracking with Wearables*. PhysioNet, 2025. [2] *SkipTrack: A Bayesian Hierarchical Model for Self-tracked Menstrual Cycle Length and Regularity*. arXiv:2508.05845, 2025. [3] *A Generative Modeling Approach to Calibrated Predictions: A Use Case on Menstrual Cycle Length Prediction*. Proceedings of Machine Learning Research (PMLR), 2026. [4] Fehring, R. J., Schneider, M., & Raviele, K. *Effectiveness of the Marquette Model of Natural Family Planning*. Journal of Obstetric, Gynecologic & Neonatal Nursing (JOGNN). [5] *Privacy and Security of Women's Reproductive Health Apps in a Changing Legal Landscape*. arXiv, 2024. [6] *Privacy Implications of FemTech Post-Roe*. Science, Technology, and Human Values, 2023. [7] *Big data research on women's health with Clue*. Clue Academic Partnerships, 2021. [8] World Health Organization (WHO). *Appropriate body-mass index for Asian populations and its implications for policy and intervention strategies*. The Lancet, 2004. (Reference for South Asian BMI cutoffs). [9] Welford, B. P. *Note on a method for calculating corrected sums of squares and products*. Technometrics, 1962. (Reference for Welford's online algorithm).