In her research work, Sahoo focuses on machine learning and data-driven modeling, with expertise in multimodal and time-series data. Her work spans mental health, smart and connected health systems and human behavior modeling, with an emphasis on interpretable and deployable AI.

Sahoo’s recent collaborative research with an interdisciplinary team from the University of Connecticut and UConn Health brings together AI, wearable computing and mental health research, with the broader goal of developing less burdensome and more continuous approaches for monitoring depression treatment progress.

“One of the challenges in mental healthcare is that understanding how a patient is doing often depends on questionnaires or clinical visits, which require patients to actively report how they feel, and these self-reports can sometimes be influenced by recall bias. Passive sensing offers a different opportunity. Everyday devices can collect information such as step count and sleep patterns with very little additional effort from the patient,” Sahoo explained.

The interdisciplinary team’s study analyzed longitudinal data from 49 patients receiving depression treatment, using daily step-count and sleep data collected through Fitbit devices. The study was conducted in a real clinical setting with participants receiving pharmacological treatment for depression. Participants were followed for up to 12 weeks, with Fitbit devices continuously collecting behavioral information. Treatment improvement was ultimately determined using clinician-administered Clinical Global Impressions assessments, rather than relying solely on self-reported outcomes.

“Our findings suggest that these everyday behavioral signals, when analyzed using advanced AI methods, may provide useful information about treatment progress. The study therefore demonstrates how computer science and AI can contribute to emerging approaches in digital health, digital phenotyping and personalized mental-health care,” Sahoo said.

To address the challenge of relatively small clinical datasets, the team developed a self-supervised contrastive learning approach using a transformer-based time-series model (PatchTST) to learn behavioral patterns from wearable data and predict whether a patient’s symptoms were improving.

From a machine learning perspective, Sahoo said the study’s goal was to investigate how much useful information could be learned from simple, passively collected signals.

“We designed the AI models to identify patterns in everyday behavioral data and examine whether those patterns could help predict treatment outcomes. An important part of the work was showing that even relatively simple signals, when analyzed appropriately using machine learning, may provide meaningful information while minimizing additional burden on patients,” said Sahoo.

The study’s results showed that combining step-count data with a one-time baseline depression assessment achieved an F1 score of up to 0.74, which increased to 0.77 when sleep data was included, Sahoo noted.

“I think this study is a good example of how computer science and AI can help us find meaningful patterns in data that are already being generated through everyday life. Changes in physical activity or sleep may appear to be simple behavioral signals, but machine learning allows us to examine how these signals change over time and how they may relate to a person’s treatment progress,” said Sahoo.

The research team’s co-authored paper, "Predicting Depression Treatment Outcome Using Daily Step Count Sensory Data," was submitted and accepted for 2026 conference presentation by the Institute of Electrical and Electronics Engineers (IEEE) and the Association for Computing Machinery (ACM). Sahoo recently presented the paper at the IEEE/ACM Conference on Connected Health: Applications, Systems and Engineering Technologies in Pittsburgh, Pennsylvania.

Sahoo said an important aspect of the research is its potential to reduce patient burden. Instead of asking patients to repeatedly complete questionnaires, wearable devices can passively collect information such as physical activity and sleep. More broadly, Sahoo said this type of study is what excites her about the potential of digital phenotyping.

“Instead of relying only on occasional snapshots of a patient’s condition, AI can potentially help us understand behavioral changes continuously and at an individual level. In the future, approaches like this could contribute to more personalized mental healthcare by helping clinicians recognize changes earlier and adapt treatment based on an individual patient’s behavioral patterns,” Sahoo said.