Embodied Music Preference Modeling: Real-Time Prediction From Wearable Gait Telemetry, Google Fit Activity, and Gesture-Based Feedback
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Music is widely deployed to enhance exercise, yet far less is known about how an exerciser’s real-time physiological state feeds back to shape musical liking. Wearable sensors now permit second-by-second coupling of gait dynamics, activity load, and affective response.
Objective
We developed a mobile workflow that predicts immediate like versus dislike judgments for unfamiliar songs by fusing clinical-grade inertial kinematics (Ambulosono), passive smartphone activity logs (Google Fit), and a single high-knees/low-knees gesture. The study tested whether momentary movement intensity biases preference, identified the strongest biometric predictors, and evaluated a sensor-aware classifier suitable for adaptive playlists.
Methods
Seventy-three healthy undergraduates performed fifty 60-second stepping-in-place trials while listening to tempo-normalised tracks spanning five genres. Ambulosono units sampled lower-limb acceleration at 200 Hz; Google Fit recorded pre- and post-trial step counts. Four gait features—mean and peak cadence, mean and peak step length—were normalised and averaged into a Composite Motivation Score (CMS). Breathlessness and fatigue ratings were logged after each track. A 500-tree random forest trained on gait variables, activity counts, perceptual deltas, and CMS classified preference using 10-fold cross-validation. Statistical tests compared liked and disliked trials for physiological change and speed strata.
Results
The protocol yielded 3 864 complete song exposures. Participants judged 54 % of tracks liked and 60 % disliked (paired t = –2.11, p = 0.036). Disliked trials exhibited larger breathlessness and fatigue increases (Wilcoxon p < 0.05). High-CMS trials showed a 68 % like-rate versus 37 % in low-CMS trials. The classifier achieved 0.78 accuracy and 0.82 AUC; permutation analysis ranked post-trial Google Fit steps, bout duration, and pre-trial steps as top predictors. Track-level analysis revealed that the ten most-disliked songs coincided with the highest mean stepping speeds, despite non-significant effects at coarse speed tiers.
Conclusions
Immediate bodily engagement and short-term physiological strain strongly colour musical appraisal. Integrating wearable kinematics, smartphone step counts, and low-friction gestures enables accurate, interpretable prediction of liking, paving the way for context-adaptive playlists and emotionally intelligent rehabilitation cues.
