AI · 4 min read · May 3, 2026
ActiNet: Self-Supervised Model Improves Wrist Activity Classification
Open-source deep learning tool outperforms random forest baselines for extracting activity intensity from wearable accelerometer data in epidemiological research.
ActiNet, a self-supervised ResNet model with HMM smoothing, classifies wrist-accelerometer activity intensity with F1 0.82, outperforming random forest baseline.
- — ActiNet combines 18-layer modified ResNet-V2 with hidden Markov model post-processing for activity classification.
- — Trained on 24-hour wrist-accelerometer data from 151 participants spanning ages 18–91.
- — Achieves macro F1 score 0.82 and Cohen's kappa 0.86 via 5-fold stratified cross-validation.
- — Outperforms established random forest + HMM baseline (F1 0.76, kappa 0.80) by 6–8 percentage points.
- — Performance gains hold consistently across age and sex subgroups without degradation.
- — Open-source tool designed for large-scale epidemiological studies linking activity to health outcomes.
- — Self-supervised pretraining reduces reliance on labeled data compared to supervised-only approaches.
Frequently asked
- ActiNet is an open-source deep learning model that classifies activity intensity from wrist-worn accelerometer data. It uses a modified ResNet-V2 neural network trained with self-supervised learning, followed by a hidden Markov model to smooth predictions. The combination achieves an F1 score of 0.82, outperforming traditional random forest approaches.