Scientific ML: Freezing-of-gait detection

The agent gets lower-back accelerometer series with expert-annotated events and builds a model that scores three freezing-of-gait event types at every timestep. It works offline and submits through a sealed leaderboard gateway.

Freezing of gait, the sudden inability to step that comes with Parkinson's disease, is dangerous, common, and hard to observe: episodes are brief and rarely happen during a clinic visit. A wearable accelerometer sees every step a patient takes. Reliable detection from that signal would move the measurement from the clinic into daily life.

The metric is mean average precision over the three event classes, computed timestep-wise, so both catching episodes and localizing them in time count. Scored submissions are quota-limited and return public scores only; the reward is the floor-normalized private score of the best-public submission.

The graded runs are hard to tell apart because they did the same work. Each trained windowed deep sequence classifiers over the raw accelerometer series, GRU-based where the logs name the architecture. Each swept window lengths from about fifteen hundred to sixteen thousand samples. Each spent the back half of the budget blending seeds and folds. The logs even overlap in what they discarded: class re-weighting and metadata features were each tried in two separate runs and dropped both times.

Most of what the blending bought was public score. One run's public number climbed submission after submission while its private score never beat the run's own first trained model. That same run caught its blend-weight search overfitting cross-validation and reverted to an earlier blend. Nothing in any log leaves this family: no other architecture, no other representation of the signal. The shared plateau therefore describes the recipe every run chose rather than a measured limit of the task.