Navigating the noise: Treatment response prediction in psychosis

Schizophrenia-spectrum disorders are highly heterogeneous, yet personalized treatment remains elusive. Antipsychotic medication is the first-line therapy for psychotic episodes, but treatment response cannot currently be predicted. Selection relies primarily on side-effect profiles, prior response, and patient preference, in the absence of validated biomarkers. This thesis advances knowledge of treatment response prediction in psychosis by reviewing the literature, developing EEG-based predictors, and critically reflecting on the feasibility of biomarker-oriented research.
A narrative review of resting-state EEG studies combining connectivity and network measures with machine-learning models showed promising results across depressive, psychotic, and trauma-related disorders. Nevertheless, the literature is constrained by small sample sizes, heterogeneous methodologies, scarce independent replication, and potential confounds such as placebo effects and natural illness course. EEG-based predictors must therefore be validated prospectively across sites before clinical use.
A PRISMA-guided systematic review of 28 fMRI studies examined functional connectivity (FC) as a biomarker for antipsychotic treatment response (AP-R). The most consistent evidence implicated FC between the striatum and the ventral attention network — particularly the putamen, caudate nucleus, insula, and anterior cingulate cortex — as a potential marker of response. Classification accuracies ranged from 50% to 93%, but external validation was limited, and methodological heterogeneity remained a key barrier to clinical translation.
In a sample of 62 first-episode psychosis (FEP) patients and 106 matched healthy controls, 60 resting-state EEG features spanning power spectrum, connectivity, and network topology were examined. No significant group differences emerged after correction for multiple testing. A random-forest regression model explained 23% of the variance in six-week positive symptom reduction. The most informative EEG features were alpha-band tree hierarchy, beta-band PLI, and delta-band betweenness centrality. These same features, combined with clinical variables, most notably hallucinatory behavior (PANSS P3), in a large antipsychotic-naïve FEP cohort explained 34% of variance , a non-significant 2% gain, suggesting clinical features are a well-suited starting point, while EEG may offer additional value in certain cases.
A reliability study in 42 healthy adults revealed that permutation entropy (PE) showed the highest test-retest stability (good-to-excellent ICC in theta and alpha bands). PLI achieved moderate-to-good reliability in the alpha band, while AECc and MST network metrics were less reliable. These findings underscore that metric reliability must be considered when selecting EEG features for biomarker models.
The thesis concludes with a theoretical reflection applying an enactive framework to EEG biomarker research. Enactivism describes psychosis as arising from dynamic brain–body–environment interactions, rather than neural processes alone. Three core assumptions of biomarker research are each shown to be only partially valid: psychosis cannot be reduced to neurobiological dysfunction; EEG captures only a slice of this interaction; and antipsychotic effects are inseparable from psychosocial context. Reliable prediction will therefore require multimodal, person-centered models integrating biological, clinical, experiential, and contextual dimensions of psychosis.