Aims: This work was done as part of the 2026 PhysioNet Challenge, which focuses on predicting cognitive impairment using multimodal sleep signals. With the proliferation of smart wearables and the development of remote photoplethysmography, continuous heart rate variability (HRV) monitoring is increasingly accessible. This study aims to develop a machine learning model utilising demographic and HRV features to detect cognitive impairment, prioritising compatibility with everyday consumer devices.
Methods: We analysed challenge electrocardiogram data, identifying R-peaks using a thresholding algorithm to extract seven standard HRV features. A Random Forest classifier was trained exclusively on these metrics and demographics (Race, Sex, Age and BMI). Separately, Mann-Whitney U tests and point-biserial correlations evaluated these features alongside an independently computed extrasystole frequency, derived via a moving mean of RR intervals.
Results: On the local cross-validation set, the Random Forest achieved an Area Under the Receiver Operating Characteristic curve (AUROC) of 0.800, an AUPRC of 0.754 and an Accuracy of 0.770. Evaluated on the hidden validation set, the AUROC remained competitive at 0.757, though AUPRC and Accuracy dropped to 0.124 and 0.058, respectively. The separate statistical analysis revealed significant correlations with the target variable for BMI (r=0.513, p<0.001), extrasystole frequency (r=-0.111, p=0.006) and pNN50 (r=-0.104, p=0.012).
Conclusion: Basic demographic and standard HRV features offer an accessible pathway for early cognitive impairment screening, evidenced by a robust AUROC, while extrasystole frequency demonstrates strong clinical relevance. Future work will integrate extrasystole metrics into the model and calibrate decision thresholds to address class imbalances and improve precision on unseen data.