Can AI accurately extract psychosis symptoms from clinical notes?

Large language models currently extract positive symptoms of schizophrenia from clinical notes with moderate accuracy (F1 scores of 0.500 to 0.647) but fail to reliably identify negative symptoms. According to empirical research synthesized by Groundwork, this disparity means AI tools are insufficient for individual clinical diagnosis but may support large-scale exploratory research. You should treat AI-extracted psychiatric data as preliminary, not definitive, for personal health decisions.

The current accuracy gap in psychiatric AI

General-purpose large language models (LLMs) demonstrate poor to moderate performance when extracting specific symptoms of schizophrenia from electronic health records. In a study evaluating 14 different LLMs across a cohort of 704 patients with severe mental illness, researchers found that macro F1 scores ranged between 0.500 and 0.647 across eight specific symptoms. This metric, which balances precision and recall, indicates that the models are correct only about half to two-thirds of the time. For context, an F1 score of 0.65 means that for every 100 symptom instances, the model accurately identifies only 65 while making errors on the remaining 35. This level of accuracy is generally considered inadequate for high-stakes individual clinical decision-making, where misdiagnosis can have severe consequences. The study explicitly notes that while proof-of-concept extractions exist for other medical conditions, the heterogeneous nature of psychiatric disorders demands stricter domain-specific validation that current models do not meet.

Positive vs. negative symptom extraction

The most critical finding is the systematic bias in how these models handle different symptom domains. Positive symptoms, such as hallucinations and delusions, were extracted with significantly higher accuracy than negative symptoms, such as flat affect or social withdrawal. This asymmetry is dangerous because it creates a distorted view of the patient's condition. If an AI tool relies on these biased extractions, it may overrepresent the more dramatic, observable symptoms while underrepresenting the subtle, internalized aspects of the disorder. This gap is not merely a technical flaw but a structural limitation in how current models are trained. They tend to latch onto explicit, keyword-heavy descriptions of acute events while missing the nuanced, often implicit language used to describe chronic negative symptoms. Until this gap is closed, any AI-derived characterization of psychotic symptoms will be systematically skewed.

Why few-shot prompting fails to fix the bias

Few-shot prompting, a common technique where examples are provided to the model to guide its output, did not significantly improve these results. The study found that adding examples of correct symptom extraction to the prompt did not narrow the performance gap between positive and negative symptoms. This suggests that the issue is not a lack of instruction but a fundamental limitation in the model's ability to understand the clinical context of negative symptoms. For patients and clinicians, this means that simply asking the AI for more context or providing more examples is not a viable solution. The underlying architecture and training data of these general-purpose models do not contain the granular, domain-specific knowledge required to distinguish between mild social withdrawal due to fatigue and the profound social anhedonia characteristic of schizophrenia. This failure highlights the need for specialized, fine-tuned models rather than relying on general-purpose LLMs for psychiatric phenotyping.

Implications for clinical research and individual care

Despite the poor individual-level accuracy, LLM-predicted positive symptoms showed comparable associations with clinical variables in regression analyses. This means that while the model may miss specific symptoms for individual patients, the aggregate data still correlates with broader clinical outcomes such as hospitalization frequency or medication adherence. This paradox makes LLMs useful for exploratory research on large cohorts but dangerous for individual care. Researchers can use these tools to identify population-level trends and risk factors, but clinicians must not use them to make treatment decisions for specific patients. The uneven attenuation of group-level associations due to individual extraction errors further complicates this picture. Some symptom domains are more affected by errors than others, meaning that even group-level analyses may be biased if not carefully adjusted for model performance. This requires rigorous statistical correction methods that go beyond standard data cleaning.

The role of biological data integration

The study suggests that combining LLM-extracted symptoms with biological data, such as genomics or neuroimaging, could enhance the utility of these tools. Biological markers may provide an independent validation of the symptoms extracted from clinical notes, helping to filter out false positives and negatives. For example, a biomarker associated with negative symptoms could help confirm the presence of these symptoms even if the LLM fails to extract them from the text. This multi-modal approach is likely the future of psychiatric AI, but it is not yet mature. Current tools are limited to text-based extraction, and the integration of biological data is still in the research phase. Patients should be cautious about any service claiming to use AI for psychiatric diagnosis without mentioning the integration of biological markers or human oversight.

Practical steps for patients and clinicians

If you or a loved one are using AI tools to manage mental health, follow these steps to ensure safety and accuracy:

1. Never use AI for diagnosis: Treat any AI-generated symptom list as a preliminary summary, not a medical opinion. Always consult a licensed psychiatrist or psychologist for diagnosis.
2. Request human review: If your healthcare provider uses AI-assisted tools, ask whether a human clinician reviews the AI's output before it is used in your care plan. Human oversight is critical to catch extraction errors.
3. Monitor for bias: Be aware that AI may overemphasize positive symptoms. If you feel your negative symptoms are being overlooked, explicitly mention them in your discussions with your clinician and in any AI-assisted tools you use.
4. Advocate for transparency: Ask your healthcare provider about the specific AI tools they use and their validation studies. Reputable providers will be able to explain the accuracy rates and limitations of their tools.
5. Combine with biological data: If possible, support the use of multi-modal approaches that combine clinical notes with biological markers. This can provide a more comprehensive and accurate picture of your condition.

Decision matrix for AI-assisted psychiatric care

The path forward for equitable AI in psychiatry

Closing the performance gap between positive and negative symptom extractions is essential groundwork to prevent systematic bias in any LLM-derived characterizations of psychotic symptoms. This requires not just better models but better training data that includes diverse, high-quality annotations of negative symptoms. It also requires rigorous validation studies that assess not just overall accuracy but the accuracy of specific symptom domains. Until these steps are taken, AI tools should be viewed as supportive, not primary, tools in psychiatric care. Patients must remain the central decision-makers in their own care, using AI as a supplementary resource rather than a diagnostic authority. The goal is not to replace clinicians but to augment their ability to manage complex, heterogeneous disorders. This requires a careful balance of technological innovation and clinical caution.

Related insights on AI in health

For a broader perspective on how AI is transforming health detection, consider our analysis on AI detecting a distributed blood metabolomic systemotype. This highlights the potential of AI in identifying subtle biological patterns that may not be apparent through traditional methods. Additionally, altered EEG patterns co-varying with specific improvements in mind health demonstrates how biological markers can provide objective evidence of mental health changes. These examples underscore the importance of combining AI with biological data to create a more comprehensive understanding of health and disease.