Leveraging machine learning to examine engagement with a digital therapeutic

Abstract

Digital Therapeutics (DTx) are evidence-based software-driven interventions for the prevention, management, and treatment of medical disorders or diseases. DTx offer the unique ability to capture rich objective data about when and how a patient engages with a treatment. Not only can one measure the quantity of patient interactions with a digital treatment with high temporal precision, but one can also assess the quality of these interactions. This is particularly useful for treatments such as cognitive interventions, where the specific manner in which a patient engages may impact likelihood of treatment success. Here, we present a technique for measuring the quality of user interactions with a digital treatment in near-real time. This approach produces evaluations at the level of a roughly four-minute gameplay session (mission). Each mission required users to engage in adaptive and personalized multitasking training. The training included simultaneous presentation of a sensory-motor navigation task and a perceptual discrimination task. We trained a machine learning model to classify user interactions with the digital treatment to determine if they were “using it as intended” or “not using it as intended” based on labeled data created by subject matter experts (SME). On a held-out test set, the classifier was able to reliably predict the SME-derived labels (Accuracy = .94; F1 Score = .94). We discuss the value of this approach and highlight exciting future directions for shared decision-making and communication between caregivers, patients and healthcare providers. Additionally, the output of this technique can be useful for clinical trials and personalized intervention.