JMIR Publications

ABSTRACT

Background:

Smartphone inertial measurement unit (IMU) and touchscreen tasks can capture Parkinson disease (PD)-related motor information, but deep-learning analyses of app-derived time-series data require participant-level leakage control and cautious clinical interpretation.

Objective:

To evaluate whether recurrent and convolutional-recurrent deep sequence models, including bidirectional long short-term memory (BiLSTM) and one-dimensional convolutional BiLSTM (Conv1D-BiLSTM) architectures, can learn PD-related information directly from smartphone-derived motor-task time series collected through CheckPD, and whether stacked ensembles improve internal case-control discrimination between neurologist-confirmed PD and neurologist-confirmed healthy-control participants.

Methods:

We performed a retrospective case-control modelling analysis of 679 participants: 344 with neurologist-confirmed PD and 335 neurologist-confirmed healthy controls. Five tasks were analysed: Rest Tremor, Postural Tremor, Balance, GaitWalk, and PinchToSize. All participant-level train, validation, and test splitting was performed before temporal windowing or touchscreen-session construction. BiLSTM and Conv1D-BiLSTM models were developed using stratified 5-fold cross-validation within the development cohort, with out-of-fold probabilities used for stacked ensembles. Final evaluation used the same randomly selected internal held-out complete-case cohort of 90 participants, balanced between 45 PD participants and 45 healthy controls with complete data for all five analysed tasks.

Results:

Conv1D-BiLSTM models generally outperformed BiLSTM models across IMU tasks, with ROC AUC increasing from 0.8087 to 0.8860 for Rest Tremor, from 0.6675 to 0.8612 for Balance, from 0.6813 to 0.8728 for Postural Tremor, and from 0.7090 to 0.8746 for GaitWalk. PinchToSize showed mixed architecture-dependent results, although both models still demonstrated useful internal case-control discrimination, with ROC AUC values of 0.8410 for BiLSTM and 0.8340 for Conv1D-BiLSTM. Among ensembles, the Conv1D-BiLSTM Main Symptoms configuration, including Rest Tremor, GaitWalk, and PinchToSize, showed the best overall discrimination and was therefore considered the leading ensemble, with the highest ROC AUC (0.9131) and PR-AUC (0.9354). The Conv1D-BiLSTM Easy configuration achieved the highest Macro F1 score (0.8534) and accuracy (85.56%) at the exported classification rule; however, these threshold-dependent metrics were considered secondary because no clinically validated operating threshold was prespecified.

Conclusions:

Deep sequence modelling of CheckPD smartphone IMU and touchscreen motor-task time series showed promising internal case-control discrimination between neurologist-confirmed PD and neurologist-confirmed healthy-control participants. Clinically, these findings suggest that brief smartphone tasks may provide objective, multidomain motor information that could complement clinical history and neurological examination in future pre-specialist assessment pathways. The results should not be interpreted as diagnostic validation or clinical deployment evidence; prospective external validation, calibration, differential-diagnosis testing, and workflow evaluation are required before patient-level use. Clinical Trial: Thai Clinical Trials Registry (TCTR20231025005) and on ClinicalTrials.gov as an observational patient registry (NCT06609681)