JMIR Publications

ABSTRACT

Background:

Passive mobile sensing provides opportunities for measuring and monitoring health status in the wild and outside of clinics. Longitudinal multimodal mobile sensor data, however, can be small, noisy, and incomplete. This makes processing, modeling, and interpretation of this data challenging. The small size of the dataset restricts them from being modeled by deep learning networks. Additionally, the current state of the art is limited to singular machine learning algorithms, making them prone to overfitting. They are constrained to forming models that are either user agnostic or personalized failing to take advantage of balancing both approaches.

Objective:

The objective of this study is to filter, rank, and output the best predictions for small multimodal longitudinal sensor data. The framework is designed to tackle datasets that are limited in size particularly targeting health studies that utilize passive multimodal sensors. The framework combines both user agnostic and personalized approaches along with a combination of ranking strategies to filter predictions.

Methods:

In this paper, we introduce a novel ranking framework to address challenges in encountered in health studies involving multimodal sensors. The framework 1) enables the processing of various combinations of sensor fusions, 2) uses user agnostic and personalized modeling approaches with appropriate cross-validation strategies, and 3) builds a tensor-based aggregation and ranking strategy for final interpretation.

Results:

The performance of the framework is validated with the help of a real dataset of adolescent diagnosed with major depressive disorder and a synthetic dataset for the prediction of change in depression of adolescent participants. Predictions output by the proposed framework achieved a 7% increase in accuracy and an increase of 13% in recall for the real dataset. The synthetic dataset saw a 5% increase in accuracy and robustness to overfitting through ensembling predictions.

Conclusions:

The framework aims to fill the gap that exists currently when modeling passive sensor data with very small number of data points. It achieves this through leveraging both user agnostic and personalized modeling techniques in tandem with an effective ranking strategy.