Psychosis is a disorder of disturbed contact with reality, characterized by hallucinations, delusions, and disorganization of speech and behavior. Recovery from psychosis is highly variable: some patients may experience a psychotic episode briefly and only once, while others suffer from recurrent episodes or chronic symptoms. The time between onset of a first psychotic episode (FEP) and start of treatment is an important and modifiable predictor of outcome.
Antipsychotics are the first-choice treatment for patients with FEP. However, in up to 40% of patients, remission is not achieved in response to the first prescribed antipsychotic drug. Non-response only becomes clear after ten weeks of treatment and is unpredictable. Clozapine is an antipsychotic with superior response rates, but is only considered after non-response to two other antipsychotics because of potential severe side-effects. Clozapine could be prescribed up to six years earlier if non-response to other antipsychotics is predicted. Development of biomarkers for response to antipsychotic treatment is therefore a crucial step towards personalized care for patients with FEP, as it will speed up remission of psychosis and prevent unnecessary side-effects in non-responders.
Recent advances in artificial intelligence in combination with suitable biomarkers now allow a personalized approach to diagnosis and treatment in psychosis patients. Electroencephalography (EEG) is a feasible and cost-effective candidate biomarker that directly measures electrophysiological brain activity with high temporal resolution. A pilot study showed that clozapine treatment response in otherwise treatment-resistant schizophrenia patients could be predicted with 85% accuracy. We have also shown the merit of machine learning of quantitative EEG characteristics in previous studies in other disorders including dementia and delirium, and found diagnostic accuracies based on this information up to 95%.
I will test the hypothesis that a combination of EEG characteristics can serve as a predictive biomarker of antipsychotic treatment response in FEP. Secondly, I will test if these EEG characteristics normalize after antipsychotic treatment for FEP as a function of treatment response.
I will use machine learning techniques, such as random forest classification, to develop a biomarker based on multiple characteristics in resting-state EEG recordings. EEG characteristics of interest include resting-state peak frequency, power spectral analysis (characterizing the frequency and amplitude of the EEG signal), frequency-dependent functional connectivity (i.e. communication between remote brain areas), and the organization of function brain networks.
EEG recordings were already obtained in 140 patients with antipsychotic drug-naive first-episode psychosis. A follow-up measurement, which consisted of clinical assessment and EEG registrations, was obtained after 6 to 12 weeks in 74 patients. Follow-up with clinical assessment without EEG registrations was available in an additional 32 patients. The random forest classifier will be used to develop a response predictor based on a set of EEG characteristics. The follow-up EEG measurement will be used to test for normalization of EEG characteristics as a function of treatment response.
The response predictor will be validated in an independent dataset of 100 patients. I will collect these data in collaboration with the Psychosis Consortium at multiple psychosis treatment centers in The Netherlands as part of the project. Measurements will be obtained at baseline and after 10 weeks of treatment.
The expected outcome of this study is an EEG-based predictor of treatment response in FEP, validated in an independent dataset. If non-response of treatment for FEP could be predicted, this could drastically speed up the start of effective treatment, and clozapine treatment could be considered directly.