Therapy dropout is often treated as a verdict on the patient: they were not ready, not motivated, or not a good fit. A new study asks a less comfortable question. What if the first conversation already contains signs of whether care will continue—and some of those signs come from the clinician?
Published on October 12 in npj Mental Health Research, the peer-reviewed study analyzed first-session transcripts from 31,146 adults using an employer-sponsored digital mental health platform. Researchers combined language-derived features with clinical, demographic, and service-use data to predict who would not attend a fourth therapy session within a year.
The scale is unusual. So is the conclusion. The transcripts added information, but the resulting model was nowhere near a dependable forecast for an individual. Its most useful lesson may be simpler than its technology: a first session should make the next step unmistakably clear.
What the model heard in 31,146 first sessions
The researchers used natural-language processing and large language models to extract features from session transcripts, including signs of patient engagement, treatment planning, therapeutic alliance, and clinician behavior. They then compared predictions made before the first appointment with predictions that also incorporated the conversation.
Before-session information produced an area under the receiver operating characteristic curve, or AUC, of 0.569. Adding first-session features raised it to 0.642. An AUC of 0.5 is no better than chance at ranking one person who drops out above one who stays; 1.0 is perfect discrimination. The improvement is real within this dataset, but 0.642 remains modest. It cannot tell a clinician with confidence what one patient will do.
That matters because dropout is common but slippery to define. A landmark meta-analysis of 669 studies and 83,834 clients estimated a weighted rate of 19.7%, while finding that the number changed with diagnosis, age, setting, clinician experience, study design, and the definition of dropout itself. In the new study, missing a fourth session within a year was the operational definition. That captures early disengagement, but it may also count people who felt better quickly, changed providers, lost benefits, faced scheduling barriers, or chose another form of care.
The clearest signal came from the clinician
Some intuitively appealing patient-level features did not perform consistently. Transcript-derived measures of motivation, ambivalence, and therapeutic alliance were not stable predictors across the analyses. The more striking association involved what clinicians did: explicitly discussing follow-up care was linked to 45% lower odds of dropout.
That figure should not be read as proof that a clear closing conversation prevents nearly half of early departures. The study was retrospective, not a randomized trial. Clinicians who discuss next steps clearly may differ in other ways; patients who are already inclined to return may invite more concrete planning. A scheduling decision recorded in speech may also reflect engagement that was already present rather than create it.
Still, the association shifts attention from trying to decode a patient's hidden commitment to examining the structure of the encounter. “What happens next?” is not administrative housekeeping. It can turn an emotionally demanding first meeting into a coherent path: when the next appointment is, what the work will focus on, and what to do if the plan feels wrong. Earlier experimental research has also suggested that setting expectations about treatment length can reduce premature termination, although that small trial of 63 clients cannot establish how the same approach would work across routine digital care.
Prediction is not prevention
The paper's limits are substantial. The data came from one commercial, employer-sponsored platform, so the findings may not generalize to public clinics, uninsured patients, in-person care, or health systems outside the United States. People had to consent to recording and transcription, which may select a group more comfortable with monitoring. Performance was weaker for non-English sessions. And language models can compress a complex exchange into categories that look precise while missing tone, culture, silence, financial pressure, or a poor match between patient and therapist.
There is also a conflict worth making explicit: the authors were employees of Spring Care, the company behind the platform, and held equity in it. That does not invalidate the analysis, but it strengthens the case for independent replication and external validation before any clinical deployment.
Transcript-based risk scores create an ethical question as well as a statistical one. Used carefully, a low-confidence flag might prompt a service to make scheduling easier or check whether the first session met the patient's needs. Used badly, it could label someone “likely to quit,” alter a clinician's expectations, or expose intimate speech to unnecessary processing. The World Health Organization's guidance on AI for health emphasizes transparency, accountability, privacy, and human oversight—requirements that become especially urgent when the raw material is a therapy conversation.
So what can the first session really predict? A little more than intake data alone, but not enough to justify automated judgments about a person. The stronger implication is practical and testable: the opening meeting is not merely diagnostic. It is a bridge to the next one, and researchers should now test whether making that bridge clearer actually helps more people continue care.
