Language-informed treatment recommendations in psychiatry: bridging the gap in individual predictability

Treatment recommendation systems in psychiatry perform suboptimally, largely due to low individual-level predictability from conventional clinical predictors. We argue that language captured through clinical interviews and electronic health records provides a rich, workflow-native substrate for estimating individualized treatment effects. We highlight opportunities for language-informed systems using modern language models and outline key challenges that must be addressed for safe and scalable deployment.