Development of a neuropsychiatric machine learning classifier to identify cases of secondary (‘organic’) psychosis

<h3>Objective</h3> A subset of patients presenting with psychosis have an underlying medical cause (‘organic’ psychosis). Early identification of these patients is essential to ensure they receive optimal care that may involve treatment of the underlying cause. If it were possible to identify which patients are at an increased likelihood of having an ‘organic’ psychosis based on their psychiatric features, they could be prioritized for further investigation to confirm the diagnosis. Using a large and representative retrospective dataset of patients diagnosed with psychosis, we aimed to develop a classification-based diagnostic model to identify cases of ‘organic’ psychosis using psychopathological features. <h3>Methods</h3> Utilising electronic health records from the largest mental health provider in the UK (South London and Maudsley NHS Foundation Trust), we identified patients with a clinical diagnosis of ‘organic’ or ‘non-organic’ psychosis between 2007 and 2022. Validated natural language processing (NLP)-based feature extraction tools were used to ascertain the presence or absence of 61 psychopathological symptoms. A weighted regularised logistic regression model was then trained to predict whether a patient ultimately received a diagnosis of an ‘organic’ or ‘non-organic’ psychotic disorder and was evaluated using 10-fold cross-validation. <h3>Results</h3> We identified 27,252 patients diagnosed with psychosis, of whom 4% (n=1,050) were assigned a diagnosis of ‘organic’ psychosis. The classifier achieved a balanced accuracy of 69% with an area under the ROC curve of 0.79. The most important features associated with ‘organic’ psychosis were waxy flexibility, concrete thinking, and visual hallucinations. In contrast, the most important variables associated with ‘non-organic’ psychosis were negative symptoms, paranoia, social withdrawal, and formal thought disorder. <h3>Conclusions</h3> In a representative sample of patients presenting with psychosis, provisional findings indicate that there may be psychopathological features that distinguish ‘organic’ psychosis. These findings, in an unprecedented sample, suggest that clinically derived prediction estimates may be useful in identifying which patients should be prioritised for further investigation to confirm or exclude a potentially reversible ‘organic’ aetiology. Expansion of the model to enhance prediction accuracy, as well as steps towards external validation to determine generalisability, will be discussed.