Two recently published papers from the lab of Tanzeem Choudhury, the Roger and Joelle Burnell Professor in Integrated Health and Technology at the Jacobs Technion-Cornell Institute at Cornell Tech, examined how smartphone data can predict patients’ own self-assessments of their condition, as well as changes in their behavior patterns in the 30 days leading to a relapse.
Early prediction of schizophrenic relapses – potentially dangerous episodes which may involve hallucinations, fears of harm, depression or withdrawal – could prevent hospitalizations, in addition to providing clinicians and patients with valuable information that could improve and personalize their care.
“The goal of this work was to predict digital indicators that are early warning signs of relapse, but these symptoms or changes can be very, very different from one individual to another,” said Dan Adler, doctoral student at Cornell Tech and first author of “Predicting Early Warning Signs of Psychotic Relapse From Passive Sensing Data: An Approach Using Encoder-Decoder Neural Networks,” which published Aug. 31 in the Journal of Medical Internet Research mHealth and uHealth.
“We tried to create an approach where we could tell a clinician: Not only is this participant experiencing unusual behavior, these are the specific things that are different in this particular patient,” Adler said. “If we can predict when someone’s symptoms are going to change before relapse, we can get them early treatment and possibly prevent an inpatient visit.”
The researchers collected smartphone data from 60 participants over one year, 18 of whom experienced relapse during that time. They used encoder-decoder neural networks – a kind of machine learning that is good at learning complex features amid highly irregular data – to detect behavior patterns such as sleep, number of missed calls, and the duration and frequency of conversations.
The method found a median 108% increase in behavior anomalies in the 30 days leading up to relapses, compared with behavior during days of relative health.
The paper used data collected in collaboration with the University of Washington, Dartmouth College and Northwell Health System. Based on the same data set, another paper – “Using Behavioral Rhythms and Multi-Task Learning to Predict Fine-Grained Symptoms of Schizophrenia,” which published Sept. 15 in Scientific Reports – used machine learning to better understand and predict symptoms from changes in behavioral rhythms passively detected by smart devices.
“We wanted to provide some actionable steps or clinically interpretable features so we can either tell the patient to take some actions or tell the clinician to suggest some early interventions,” said Cornell Tech doctoral student Vincent Tseng, the Scientific Reports paper’s co-first author.
Co-first author is Akane Sano, assistant professor in the Department of Electrical Engineering and Computer Science at Rice University and formerly a visiting scientist in Choudhury’s People Aware Computing Lab.
That study explored the relationship between symptom conditions and behavioral rhythms over periods that were less than, greater than or equal to 24 hours – each of which might influence mental disorders differently. They then analyzed the data using multitask learning – a way of training machine learning models to predict multiple related tasks simultaneously, while accounting for the tasks’ similarities and differences – to predict participants’ scores on self-assessments for 10 different symptoms, such as feeling depressed or hearing voices.
They found that their model was significantly better at predicting patients’ self-assessments than existing models. Different types of rhythms also impacted different symptoms, they found – circadian, or daily, rhythms had an influence on sleep, feeling calm and feeling social; while ultradian rhythms, of less than a day, impacted seeing things and hearing voices.
“Taken together, these different types of rhythms provide a more intuitive way to interpret the relationship between a patient’s behaviors and their symptoms,” the authors wrote. “This can determine when and the type of intervention to be delivered to avoid certain symptoms or prevent them from worsening.”
For example, if the system notices a change in the amount of ambient noise over a few hours – which was shown to affect hallucinations – the person could be prompted to move to a quieter spot.
The findings could not only reduce dangerous episodes and cut health care costs for patients with schizophrenia, but could be adapted for other mental health disorders, such as depression, Adler said. Extreme behavioral changes precede symptoms worsening across different mental health conditions, and the same prediction approach could be applied to these different conditions.
“By focusing on changes in behavioral routines and misalignment with underlying biological rhythms, we expect our approach to generate clinically actionable insights that generalize across a diverse demographic of users,” Choudhury said.
Co-authors of the mHealth paper include researchers at Dartmouth College and the Dartmouth Geisel School of Medicine, the University of Washington, Northwell Health and the Vanguard Research Group. The Scientific Reports paper was co-authored with researchers from Dartmouth, the Dartmouth Geisel School of Medicine, the University of Washington, Vanguard Research Group and Facebook.
Both papers were supported by the National Institutes of Health’s Exceptional Unconventional Research Enabling Knowledge Acceleration (EUREKA) program.