Speech monologues recorded in naturalistic settings provide opportunities to characterize mental illness phenomenology and detect symptom exacerbation. Large language models (LLMs) offer new possibilities for automating this process, as they require annotated data primarily for evaluation rather than training. In this paper, we present a novel automated, multi-agent LLM pipeline for the fine-grained, multi-label extraction of language suggestive of delusional beliefs, associated affective responses, and behavioral responses from transcripts of naturalistic audio diaries collected from people with moderate persecutory ideation. Evaluating an ensemble of three foundation models, we demonstrate that detailed diagnostic prompt instructions successfully reduce false positives for delusional theme classification, but also constrain the interpretation of affective or behavioral responses. Furthermore, comparing multi-agent adjudication frameworks reveals a critical divergence from standard NLP benchmarks: complex conversational debate between agents diminishes accuracy on clinically ambiguous text by inducing premature consensus. Instead, majority voting establishes robust performance (Micro F1 of 0.872 and 0.779 for delusion detection and classification respectively). This work provides a validated and scalable pipeline for the automated detection and characterization of content suggesting delusional beliefs in naturalistic speech.

A growing amount of psychiatric research incorporates machine learning and natural language processing methods, however findings have yet to be translated into actual clinical decision support systems. Many of these studies are based on relatively small datasets in homogeneous populations, which has the associated risk that the models may not perform adequately on new data in real clinical practice. The nature of serious mental illness is that it is hard to define, hard to capture, and requires frequent monitoring, which leads to imperfect data where attribute and class noise are common. With the goal of an effective AI-mediated clinical decision support system, there must be computational safeguards placed on the models used in order to avoid spurious predictions and thus allow humans to review data in the settings where models are unstable or bound not to generalize. This paper describes two approaches to implementing safeguards: (1) the determination of cases in which models are unstable by means of attribute and class based outlier detection and (2) finding the extent to which models show inductive bias. These safeguards are illustrated in the automated scoring of a story recall task via natural language processing methods. With the integration of human-in-the-loop machine learning in the clinical implementation process, incorporating safeguards such as these into the models will offer patients increased protection from spurious predictions.