LifeMH-FL: A Lifecycle-Stratified Mental Health Corpus for Federated Learning

LifeMH-FL is the first lifecycle-stratified mental health corpus designed for federated learning research. It contains 111,191 dual-labelled records partitioned across three women's health lifecycle cohorts: S1 — Menstrual (77,259 records): sourced from MENST (clinical Q&A on menstrual disorders) S2 — Perinatal (31,743 records): sourced from Dreaddit (Reddit stress narratives, keyword-filtered) S3 — Menopausal (2,189 records): sourced from Women Health Mini (patient–provider dialogue) Each record carries two labels: (1) mental health condition — Depression, Anxiety, Stress, Suicidal, Normal (per DSM-5); and (2) risk severity — Low, Medium, High. Labels are pseudo-annotations generated by a locally-deployed Meta-Llama-3.1-8B-Instruct to preserve privacy; out-of-vocabulary outputs are normalised to the nearest DSM-5 class. Construction pipeline: Records are partitioned via a 33-term lifecycle keyword lexicon (word-boundary regex; unmatched records discarded), de-duplicated by MD5 hashing, filtered to ≥10 characters, and label-normalised. A normalization_report.json documents all label corrections applied. Novelty: LifeMH-FL is the first corpus to (a) stratify women's mental health text by biological lifecycle stage, (b) provide simultaneous condition and risk-severity labels, and (c) be explicitly structured as non-IID federated silos with a 35× volume disparity between S1 and S3 — reflecting real-world data scarcity in underserved populations. The extreme class imbalance (ratio 28.38; Suicidal class = 1.28%) mirrors clinical reality and makes the corpus a challenging benchmark for cost-sensitive and federated learning methods. Intended use: Research into privacy-preserving federated LLM fine-tuning, mental health NLP, and lifecycle-aware machine learning. Labels are for model training and evaluation only and do not constitute clinical diagnoses. Ethics: All source datasets are publicly available under open research licenses. No primary human subject data was collected and no IRB approval was required. Dataset and code: https://github.com/sayoojd/FedlifeLLM DOI of Paper: 10.1109/TCE.2026.3711046