Hipak Virtual 2021

Preliminary study offers machine learning solution for ADHD treatment monitoring

לילך שמר-מאירי 1 Aharon Schif 2
11. The Pediatric Neurology and Child developmental unit, Carmel medical center and Clalit health services, ישראל
2Child developmental unit, Clalit health services, ישראל

Objectives: ADHD medical treatment monitoring processes, currently rely on subjective parameters such as patient, family and educators’ reports. This may lead to a highly inefficient process that leads to unsuccessful results, based on possible wrong perceptions of educators and families regarding medication effectiveness and treatment success. Our study was aimed at examining changes in physiological markers in ADHD patients during titration, and associating them with different features characterizing the medical treatment. Such associations might enable the development of a machine learning based platform to facilitate decision making during treatment.

Method: The study included 37 male boys aged 7 to 12 years, underwent clinical diagnosis by physicians and were referred to additional diagnosis using TOVA computerized test. Physiological parameters of the participants were measured using the Empatica E4 wearable watch (“Iluria Ltd”) that provides readings of several markers: acceleration, angular acceleration, heart rate, heart rate variability, inter beat interval, galvanic skin response and skin temperature. Measurements for each participant were taken twice, each time for 20 minutes during a TOVA test. The first TOVA test was taken before medication has started, and the second TOVA test was taken 90 minutes after taking 10 mg of Ritalin. Regression ensemble model using LSBoost was used to regress the differences in TOVA scores before and after medication on the physiological parameters.

Results: The performance of the classifier that was designed to predict whether patients are on medication based on their physiological parameters was evaluated using a leave-one-out procedure. This resulted in 72% precision on average. P-value for the model of each patient was computed using a permutation test, yielding significant p-values for 25 out of the 37 patients.

Conclusions: We proved the concept that the use of machine learning allows for linking changes in physiological markers to various features of medical treatment of ADHD patients (further development and enhancement of machine learning models and algorithms is assumed to improve precision further). These links might be used to further develop a biomarker-based platform that can continuously and passively monitor patients and facilitate medical decision making. Such a platform is expected to reduce trial and error, and to optimize drug usage during treatment.