Forecasting emotional disorder symptoms: A dynamical systems and time-series machine learning approach - PROJECT SUMMARY Emotional disorders are among the primary causes of disability and impairment worldwide. Predicting their future course can be used to improve disease management and to deliver more timely and individually tailored treatment. However, accurate forecasting is notoriously difficult. A major challenge is that the static features traditionally used for prediction (i.e., cross-sectional symptom data) are not theoretically optimal for forecasting disease progression. Overwhelming evidence demonstrates that emotional disorders are characterized by behaviors that unfold dynamically over time. A more principled disorder conceptualization is needed that emphasizes dynamic time-series features of emotional psychopathology. To address this gap, dynamical systems and complex network theory have been proposed as theoretical models that can explain the dynamic behavior of emotional disorders. In the current proposal, we will create a precision-medicine framework for predicting emotional disorder symptoms that leverages theoretically relevant time-series measures using modern “forecasting” machine learning methods. Patients with emotional disorders will be recruited to undergo a 4-week ecological momentary assessment period during which symptom and affect data will be collected 5 times per day and passive sensor data will be continuously monitored. Aim 1 will evaluate the predictive utility of “forecasting” machine learning using EMA-based dynamical systems and network features in predicting future transdiagnostic symptom domains. Aim 2 will evaluate the added predictive utility of incorporating dynamic time- series features from passive sensor data alongside EMA data in predicting future transdiagnostic symptom domains. In Aim 3, we will develop dynamic time-series phenotypes of emotional disorders using both self- reported EMA and sensor data. The current proposal will result in the first forecasting machine learning framework designed to leverage temporally dynamic features for predicting the future course of symptoms, thereby better capturing the dynamic nature of emotional psychopathology. This is significant as it facilitates more accurate monitoring and just-in-time interventions for emotional disorders, alleviating one of the major causes of disease and burden worldwide. Furthermore, a tailored training plan was created to cultivate competencies in advanced time-series computational techniques and career development, including expertise in (1) early warning signals in dynamical systems and network theory, (2) time-series machine learning (forecasting), (3) unsupervised machine learning, (4) passive sensor data, and (5) grantsmanship and transition to academic independence, with guidance from Dr. Paola Pedrelli (primary mentor), Dr. Richard McNally (co- mentor), Dr. Jordan Smoller (co-mentor), and Drs. Rosalind Picard and Amanda Baker (consultants). Research and training activities will harness the exceptional resources at Massachusetts General Hospital/Harvard Medical School. This training plan will facilitate a future research program leveraging computational modeling to advance precision medicine and personalized care for emotional disorders.
