The case for prediction in depression care

Trial and error is the default

Depression is one of the leading causes of disability worldwide. Effective treatments exist — antidepressants across multiple classes, several forms of psychotherapy, neuromodulation, and other interventional approaches. The problem isn't a lack of options. It's that we have no systematic way to determine which option is right for a given person.

Current practice follows stepwise protocols: start with a first-line treatment, wait six to eight weeks to evaluate response, and if it doesn't work, try the next option. For many people, this process repeats multiple times before something works — if they don't disengage from care entirely.

The human cost is measured in suffering and, in all too many cases, death. The system cost is measured in billions spent on treatments that don't work.

Why prediction is hard

Predicting treatment response requires data that largely doesn't exist at the right scale and detail. Academic datasets and clinical trials have established that prediction is possible in principle — and Eiro's roots in the Wellcome Leap MCPsych program gave us direct access to this foundational work. The problem is that these datasets, while valuable for research, are too small and too narrow to power a safe, reliable, real-world product.

Clinical trials are designed to test whether a treatment works on average — not to identify who it works for. Electronic health records capture prescriptions and diagnoses but miss the patient experience: side effects, functional improvement, quality of life, the lived reality of treatment. What's needed is a dataset that connects the full picture — clinical history, treatment decisions, patient-reported outcomes, and longitudinal follow-up over months — at a scale that supports real clinical use.

From research to real-world data

Eiro has been working on this problem for over four years, starting within the Wellcome Leap MCPsych program at Vanderbilt. That work produced foundational models and access to academic datasets that demonstrate the viability of treatment prediction. The science works. The challenge now is data at scale.

Eiro Engage is how we bridge that gap. By compensating people directly to document their treatment experiences, we're building a large-scale dataset of real depression treatment journeys — the kind of data that academic studies can't produce and electronic health records don't capture. This dataset, combined with what we've already built, becomes the foundation for prediction models reliable enough to deploy in clinical settings.

What we mean by prediction

Eiro is building models that take a patient's clinical profile — diagnosis details, treatment history, comorbidities, demographics, patient-reported measures — and output a probability estimate for response to each available treatment option.

This is not a replacement for clinical judgment. It's a tool to augment it. A clinician considering two treatment paths for a patient would have access to data-driven estimates of which is more likely to work, informed by outcomes from thousands of similar patients.

The accuracy and reliability of these models is entirely dependent on the quality and scale of the underlying data — which is why the data collection phase is not a preliminary step. It is the work.