RESUMEN
Variational inference provides a way to approximate probability densities through optimization. It does so by optimizing an upper or a lower bound of the likelihood of the observed data (the evidence). The classic variational inference approach suggests maximizing the Evidence Lower Bound (ELBO). Recent studies proposed to optimize the variational Rényi bound (VR) and the χ upper bound. However, these estimates, which are based on the Monte Carlo (MC) approximation, either underestimate the bound or exhibit a high variance. In this work, we introduce a new upper bound, termed the Variational Rényi Log Upper bound (VRLU), which is based on the existing VR bound. In contrast to the existing VR bound, the MC approximation of the VRLU bound maintains the upper bound property. Furthermore, we devise a (sandwiched) upper-lower bound variational inference method, termed the Variational Rényi Sandwich (VRS), to jointly optimize the upper and lower bounds. We present a set of experiments, designed to evaluate the new VRLU bound and to compare the VRS method with the classic Variational Autoencoder (VAE) and the VR methods. Next, we apply the VRS approximation to the Multiple-Source Adaptation problem (MSA). MSA is a real-world scenario where data are collected from multiple sources that differ from one another by their probability distribution over the input space. The main aim is to combine fairly accurate predictive models from these sources and create an accurate model for new, mixed target domains. However, many domain adaptation methods assume prior knowledge of the data distribution in the source domains. In this work, we apply the suggested VRS density estimate to the Multiple-Source Adaptation problem (MSA) and show, both theoretically and empirically, that it provides tighter error bounds and improved performance, compared to leading MSA methods.
RESUMEN
BACKGROUND: Suicide is the second leading cause of death in adolescents, and self-harm is one of the strongest predictors of death by suicide. The rates of adolescents presenting to emergency departments (EDs) for suicidal thoughts and behaviors (STBs) have increased. Still, existing follow-up after ED discharge is inadequate, leaving a high-risk period for reattempts and suicide. There is a need for innovative evaluation of imminent suicide risk factors in these patients, focusing on continuous real-time evaluations with low assessment burden and minimal reliance on patient disclosure of suicidal intent. OBJECTIVE: This study examines prospective longitudinal associations between observed real-time mobile passive sensing, including communication and activity patterns, and clinical and self-reported assessments of STB over 6 months. METHODS: This study will include 90 adolescents recruited on their first outpatient clinic visit following their discharge from the ED due to a recent STB. Participants will complete brief weekly assessments and be monitored continuously for their mobile app usage, including mobility, activity, and communication patterns, over 6 months using the iFeel research app. Participants will complete 4 in-person visits for clinical assessment at baseline and at the 1-, 3-, and 6-month follow-ups. The digital data will be processed, involving feature extraction, scaling, selection, and dimensionality reduction. Passive monitoring data will be analyzed using both classical machine learning models and deep learning models to identify proximal associations between real-time observed communication, activity patterns, and STB. The data will be split into a training and validation data set, and predictions will be matched against the clinical evaluations and self-reported STB events (ie, labels). To use both labeled and unlabeled digital data (ie, passively collected), we will use semisupervised methods in conjunction with a novel method that is based on anomaly detection notions. RESULTS: Participant recruitment and follow-up started in February 2021 and are expected to be completed by 2024. We expect to find prospective proximal associations between mobile sensor communication, activity data, and STB outcomes. We will test predictive models for suicidal behaviors among high-risk adolescents. CONCLUSIONS: Developing digital markers of STB in a real-world sample of high-risk adolescents presenting to ED can inform different interventions and provide an objective means to assess the risk of suicidal behaviors. The results of this study will be the first step toward large-scale validation that may lead to suicide risk measures that aid psychiatric follow-up, decision-making, and targeted treatments. This novel assessment could facilitate timely identification and intervention to save young people's lives. INTERNATIONAL REGISTERED REPORT IDENTIFIER (IRRID): DERR1-10.2196/46464.
RESUMEN
Previous studies have demonstrated the feasibility and promise of wearable sensors as objective measures of motor impairment in Parkinson disease and essential tremor. However, there are few published studies that have examined such an application in Huntington disease (HD). This report provides an evaluation of the potential to objectively quantify chorea in HD patients using wearable sensor data. Data were derived from a substudy of the phase 2 Open-PRIDE-HD study, where 17 patients were screened and 15 patients enrolled in the substudy and ultimately 10 patients provided sufficient wearable sensor data. The substudy was designed to provide high-resolution data to inform design of predictive algorithms for chorea quantification. During the entire course of the 6-month study, in addition to chorea ratings from 18 in-clinic assessments, 890 home assessments, and 1,388 responses to daily reminders, 33,000 h of high-resolution accelerometer data were captured continuously from wearable smartwatches and smartphones. Despite its limited sample size, our study demonstrates that arm chorea can be characterized using accelerometer data during static assessments. Nonetheless, the small sample size limits the generalizability of the model. The sensor-based model can quantify the chorea level with high correlation to the chorea severity reported by both clinicians and patients. In addition, our analysis shows that the chorea digital signature varies between patients. This work suggests that digital wearable sensors have the potential to support clinical development of medications in patients with movement disorders, such as chorea. However, additional data would be needed from a larger number of HD patients with a full range of chorea severity (none to severe) with and without intervention to validate this potentially predictive technology.
RESUMEN
BACKGROUND: A growing number of clinical trials use various sensors and smartphone applications to collect data outside of the clinic or hospital, raising the question to what extent patients comply with the unique requirements of remote study protocols. Compliance is particularly important in conditions where patients are motorically and cognitively impaired. Here, we sought to understand patient compliance in digital trials of two such pathologies, Parkinson's disease (PD) and Huntington disease (HD). METHODS: Patient compliance was assessed in two remote, six-month clinical trials of PD (n = 51, Clinician Input Study funded by the Michael J. Fox Foundation for Parkinson's Research) and HD (n = 17, sponsored by Teva Pharmaceuticals). We monitored four compliance metrics specific to remote studies: smartphone app-based medication reporting, app-based symptoms reporting, the duration of smartwatch data streaming except while charging, and the performance of structured motor tasks at home. RESULTS: While compliance over time differed between the PD and HD studies, both studies maintained high compliance levels for their entire six month duration. None (- 1%) to a 30% reduction in compliance rate was registered for HD patients, and a reduction of 34 to 53% was registered for the PD study. Both studies exhibited marked changes in compliance rates during the initial days of enrollment. Interestingly, daily smartwatch data streaming patterns were similar, peaking around noon, dropping sharply in the late evening hours around 8 pm, and having a mean of 8.6 daily streaming hours for the PD study and 10.5 h for the HD study. Individual patients tended to have either high or low compliance across all compliance metrics as measured by pairwise correlation. Encouragingly, predefined schedules and app-based reminders fulfilled their intended effect on the timing of medication intake reporting and performance of structured motor tasks at home. CONCLUSIONS: Our findings suggest that maintaining compliance over long durations is feasible, promote the use of predefined app-based reminders, and highlight the importance of patient selection as highly compliant patients typically have a higher adherence rate across the different aspects of the protocol. Overall, these data can serve as a reference point for the design of upcoming remote digital studies. TRIAL REGISTRATION: Trials described in this study include a sub-study of the Open PRIDE-HD Huntington's disease study (TV7820-CNS-20016), which was registered on July 7th, 2015, sponsored by Teva Pharmaceuticals Ltd., and registered on Clinicaltrials.gov as NCT02494778 and EudraCT as 2015-000904-24 .