Joshua Curtiss, an assistant professor of applied psychology at Northeastern University in Boston, is training machine-learning models to do something meteorologists and economists have long attempted: predict what comes next. But instead of forecasting storms or stock prices, Curtiss and his team are trying to predict human emotions—a development that could reshape how mental health support is delivered.
The question driving this work is deceptively simple: Can we see patterns in how someone feels today that tell us how they'll feel tomorrow? For people living with anxiety, depression, or other emotional disorders, having that advance notice could be transformative. Rather than waiting for a crisis or struggling through treatment that hasn't been tailored to their specific needs, individuals could receive proactive, personalized interventions based on what their emotional trajectory is likely to be.
Curtiss leads the Computational Clinical Psychology Lab and has long observed a fundamental problem in how mental health care operates. "Oftentimes, the treatment or intervention for these disorders, which can include therapy or medication, are a one-size-fits-all approach," he explained. What works for one person may leave another unchanged. Humans, he argues, are too diverse for cookie-cutter solutions.
To test whether emotion prediction was feasible, Curtiss conducted a pilot study with 34 people who had been officially diagnosed with an emotional disorder. Five times each day over two weeks, participants rated their emotions on a seven-point scale, reporting on four specific feelings: contentedness, cheerfulness, sadness, and anxiousness. That daily logging—140 data points per person over the study period—became the fuel for machine-learning models ranging from simple averages to sophisticated neural networks that mimic how the brain processes information.
The results were promising. The individual models proved more accurate than a baseline comparison in predicting someone's emotions about one day in the future. But the findings revealed something equally important: different models worked better for different people and different emotions. For contentedness and cheerfulness, a model forecasting based on past performance was most accurate. For sadness and anxiousness, an ensemble model that blended results from multiple approaches performed best. This variation itself is the point—it underscores why personalization matters so much.
Though this research remains in early "proof of concept" stages, the implications ripple outward. The simplest application would be giving someone a heads-up: based on our analysis of your patterns, you might feel more anxious on Thursday, so here's what you can do to prepare. More ambitiously, the predictions could empower people to shift their behavior—adopting a habit or avoiding a trigger—to preempt emotional difficulties before they take hold. "It could give some bandwidth and wiggle room to preempt or offset some of the things we think could be happening in your future," Curtiss said.
Don Robinaugh, an assistant professor in Northeastern's Applied Psychology department who was not involved in the study, captured the broader significance: "Better prediction really means better care." With more validation, Curtiss believes forecasting machine learning could open a new era of mental health support—one built around the individual rather than around a standard protocol. The goal is not to predict emotions perfectly, but to know each person well enough to help them thrive.