FairTune: A Bias-Aware Fine-Tuning Framework Towards Fair Heart Rate Prediction from PPG

FairTune is a bias-aware fine-tuning recipe for PPG-GPT, not a new pretrained encoder. The authors start from a PPG-GPT checkpoint pretrained on UCSF ICU data (over 200 million segments; four sizes from 19M to 1B) and fully fine-tune it for heart-rate regression. They resample PPG-DaLiA (wrist, 64 Hz), BUT-PPG (Xiaomi Mi9 camera, 30 Hz), and a MIMIC ICU set at 125 Hz to 40 Hz and use 1-second windows. The MIMIC extract is cited via Charlton et al. 2022 and is not tied to one hub dataset page. Table 2 reports unbalanced fine-tuning MAE 17.62 bpm on DaLiA, down from a pretrained MAE 89.54 bpm (80%). The same table reports MAE 19.23 bpm on BUT-PPG, down from 78.22 bpm (75%). It reports MAE 24.73 bpm on MIMIC, down from 93.81 bpm (74%). Same-dataset Table A1 unbalanced MAE is 4.28 bpm on DaLiA, 9.94 bpm on BUT-PPG, and 9.76 bpm on MIMIC. The gender gap is the absolute male–female MAE difference. Table 3 does not shrink it on every transfer: DaLiA to MIMIC moves from 0.80 bpm unbalanced to 1.06 bpm with inverse-frequency weighting, 1.20 bpm with GroupDRO, and 1.16 bpm with adversarial debiasing, while DaLiA to BUT-PPG GroupDRO lowers the gap from 2.36 to 0.68 bpm. The PDF uses a Proceedings of Machine Learning Research class header and names no volume. No code URL is given.

• Authors: Panchumarthi, Kataria, Wu, Hu, Fedorov, Kwak (Emory; Wu also University of Oklahoma)
• arXiv:2509.16491v1 (20 Sep 2025); DOI 10.48550/arXiv.2509.16491; PDF metadata licence CC BY 4.0; no code URL; PMLR class header, no volume
• Full fine-tune of PPG-GPT. Reported results use 19M; the paper also scales from 19M to 1B. Checkpoint pretrained on UCSF ICU data, over 200M segments
• Table 2 unbalanced MAE: DaLiA 89.54 to 17.62 bpm (80%), BUT-PPG 78.22 to 19.23 bpm (75%), MIMIC 93.81 to 24.73 bpm (74%). Table A1 same-dataset unbalanced MAE: 4.28, 9.94, 9.76 bpm
• Table 3, 19M medians, DaLiA to BUT-PPG: unbalanced MAE 19.23 gap 2.36; IF 13.44 / 1.30; GroupDRO 15.47 / 0.68; ADV 14.49 / 0.70 bpm
• Table 3 does not shrink every gap. BUT-PPG to DaLiA: unbalanced 6.99 vs IF 6.86 bpm. DaLiA to MIMIC: unbalanced 0.80 vs IF 1.06, GroupDRO 1.20, ADV 1.16 bpm. At 1B, DaLiA to MIMIC is MAE 26.66 gap 1.27 (19M: 24.74 and 0.80); MIMIC to BUT-PPG is MAE 16.23 gap 2.65 (19M: 18.68 and 4.93)

Foundation models pretrained on physiological data such as photoplethysmography (PPG) signals are increasingly used to improve heart rate (HR) prediction across diverse settings. Fine-tuning these models for local deployment is often seen as a practical and scalable strategy. However, its impact on demographic fairness—particularly under domain shifts—remains underexplored. We fine-tune PPG-GPT — a transformer-based foundation model pretrained on intensive care unit (ICU) data — across three heterogeneous datasets (ICU, wearable, smartphone) and systematically evaluate the effects on HR prediction accuracy and gender fairness. While fine-tuning substantially reduces mean absolute error (up to 80%), it can simultaneously widen fairness gaps, especially in larger models and under significant distributional characteristics shifts. To address this, we introduce FairTune, a bias-aware fine-tuning framework in which we benchmark three mitigation strategies: class weighting based on inverse group frequency (IF), Group Distributionally Robust Optimization (GroupDRO), and adversarial debiasing (ADV). We find that IF and GroupDRO significantly reduce fairness gaps without compromising accuracy, with effectiveness varying by deployment domain. Representation analyses further reveal that mitigation techniques reshape internal embeddings to reduce demographic clustering. Our findings highlight that fairness does not emerge as a natural byproduct of fine-tuning and that explicit mitigation is essential for equitable deployment of physiological foundation models.

Verbatim from arXiv (opens in a new tab) · arXiv metadata (abstracts are released under CC0)

No results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.

• PPG-DaLiAPPG dataset for motion compensation / HR15 subjects · PPG · ECG · IMU
• Brno University of Technology Smartphone PPG Database (BUT PPG)50 subjects · PPG · ECG

• FairTunetransformerBias-aware fine-tuning of PPG-GPT for heart-rate prediction across ICU, wrist, and smartphone PPGPPG2025
• GPT-PPGtransformerFoundation modelPPG2025