A routine doctor visit may become an early warning layer for depression

Analyzing 1,108 audio-recorded primary care chats from the Establishing Focus study, researchers trained three supervised models—Sentence-BERT+LR, LIWC+LR, and ModernBERT—plus a zero-shot GPT-OSS—to spot depression using PHQ-9 labels. It turns out the best performer wasn’t one of the fine-tuned heirlooms but the open-weight newcomer GPT-OSS, clocking an AUROC of 0.774 and AUPRC of 0.510. That’s respectable for clinical decision support, but still shy of the 0.9+ AUROC typically demanded for screening tools. Still, in a domain where depression is routinely missed, even a marginal gains can shift outcomes. Primary care remains a pressure cooker for underdiagnosis, and natural language models now offer a chance to triangulate risk from patient-doctor dialogue without adding more forms to the EHR stack.

In practice, digital scribing platforms—already logging these encounters—could embed lightweight speech-to-text pipelines upstream of clinical review, turning idle transcripts into early alerts. LIWC+LR kept pace with a 0.742 AUROC using only lexical hand-crafted features, an encouraging sign that simpler architectures can extract meaningful signal from routine chatter. Yet the real chasm isn’t algorithmic performance; it’s integration friction and clinician trust—two variables rarely optimized in academic bake-offs.

Zero-shot models like GPT-OSS sidestep expensive labeled datasets, but their edge here is modest. The 0.774 AUROC places this approach in the same league as earlier deep-learning trials on clinical text, suggesting incremental progress rather than a leap. Benchmarks matter, yet real-world deployment demands calibration across dialects, accents, and clinic workflows. Players note that primary care visits average seven minutes; any tool that disrupts flow will be shelved. Companies eyeing this opportunity should focus less on headline metrics and more on seamless backend integration, secure storage, and explainable outputs that clinicians can override without friction.

The study’s setting—routine primary care—also signals a pivot toward embedded AI, where tools earn their keep by quietly augmenting existing tools rather than launching new ones. It’s not about replacing psychiatrists; it’s about catching the silent majority slipping through 15-minute slots. If GPT-OSS can reach 0.85 AUROC with modest fine-tuning on dialect-diverse corpora, the gap between demo and clinic may finally shrink.