Explainable Machine Learning for Physical Activity Intensity Classification using Multimodal Wearable Sensors and Sensor Ablation Analysis

Abstract

Wearable sensors are used to assess whether individuals reach moderate or vigorous physical activity, but many classification studies test models on people who also supplied training data and report accuracy without any statement of uncertainty. This study classified activity intensity in people withheld from training and asked which sensors are needed and how reliably predictions generalise to new participants. Previously collected recordings of twelve activities from eight adults wearing wrist, chest and ankle inertial sensors and a heart-rate monitor were labelled low, moderate or vigorous using the 2024 Adult Compendium of Physical Activities. Logistic regression, random forest and gradient boosting were evaluated by leave-one-participant-out validation with all tuning nested inside the training participants, across eleven sensor combinations. A heart-rate normalisation that needs no labelled calibration session was compared with raw heart rate, and prediction sets were built with conformal methods calibrated on unseen participants. With all sensors, random forest achieved a macro-F1 score of 0.954 and gradient boosting achieved 0.951. A wrist sensor alone reached 0.87 and adding heart rate raised this to 0.92, an improvement seen in seven of eight participants. The personalised heart-rate measure gave no clear gain over raw heart rate. Prediction sets calibrated using held-out participants achieved approximately 90% marginal coverage, whereas sets calibrated on training data covered only 73% of cases. Vigorous activity was under-covered until coverage was calibrated separately for each intensity class, which raised it from 77% to 90%. A wrist sensor combined with heart-rate monitoring is a promising configuration for further evaluation, and stated uncertainty should be calibrated on people outside the training set.