Subtyping first-episode psychosis based on longitudinal symptom trajectories using machine learning

Abstract

Clinical course after first episode psychosis (FEP) is heterogeneous. Subgrouping and predicting longitudinal symptom trajectories after FEP may help develop personalized treatment approaches. We utilized k-means clustering to identify clusters of 411 FEP patients based on longitudinal positive and negative symptoms. Three clusters were identified. Cluster 1 exhibits lower positive and negative symptoms (LS), lower antipsychotic dose, and relatively higher affective psychosis; Cluster 2 shows lower positive symptoms, persistent negative symptoms (LPPN), and intermediate antipsychotic doses; Cluster 3 presents persistently high levels of both positive and negative symptoms (PPNS), and higher antipsychotic doses. We predicted cluster membership (AUC of 0.74) using ridge logistic regression on baseline data. Key predictors included lower levels of apathy, affective flattening, and anhedonia/asociality in the LS cluster, compared to the LPPN cluster. Hallucination severity, positive thought disorder and manic hostility predicted PPNS. These results help parse the FEP trajectory heterogeneity and may facilitate the development of personalized treatments.

Fig. 1. Optimal cluster number and the corresponding cluster size.

a Clustering performance by number of clusters. X axis is the number of clusters, y axis is the ratio of inertia to silhouette score, which reflects the compactness within each cluster and the separation between clusters. The optimal number of clusters is 3, when the ratio of inertia to silhouette score reaches its minimum. b Cluster size. The number of patients in each cluster and their proportions. The trajectories these clusters represent are discussed below.

Fig. 2. Symptom score trajectories.

a Overall symptom trajectories with mean value and standard error. X axis is timepoint in months, y axis is the score of positive or negative symptom measured by SAPS or SANS. Solid line = negative symptoms; dashed line = positive symptoms. b Relative change from baseline. Each mean value was subtracted from the mean value at baseline to illustrate the relative change at each timepoint. c Positive symptom score trajectories with Kruskal-Wallis test significant levels, * means p < 0.05, ** means p < 0.01, *** means p < 0.001 after bonferroni correction. d Negative symptom score trajectories, same as positive symptom. Color coding: green for cluster 1 LS, blue for cluster 2 LPPN, red for cluster 3 PPNS.

Fig. 3. Diagnoses at baseline.

Normalized bar plot of schizophrenia spectrum and affective psychosis for three clusters (One patient in cluster 1 with substance-induced psychosis was excluded from this figure).

Fig. 4. Cluster Membership Prediction and import features for predicting each cluster.

a Permutation test. Histogram is the distribution of AUC for random prediction. Red vertical line is the actual prediction AUC. b–d The 15 most important features in the prediction for each cluster. X axis is SHAP value, representing the impact on model output of a specific feature, y axis is the feature name. Red indicates a higher value of the feature, and blue indicates a lower value.

Fig. 5. Treatment and adherence trajectories.

a Antipsychotic dosage trajectories. X axis is timpoint in months, y axis is the converted doses measured by chlorpromazine equivalents. Kruskal-Wallis test was conducted to evaluate the difference among three clusters.* notes significant levels, * means p < 0.05, ** means p < 0.01, *** means p < 0.001 after bonferroni correction. b Mean adherence trajectories with standard error. Kruskal-Wallis test was also conducted at each timepoint, but none is significant. Adherence levels of 0%, 25%, 50%, 75%, and 100% means never, very infrequently, sometimes, quite often, and always taking medicines. Black dashed line is the mean adherence level for all patients.