There is broad agreement among psychiatrists, based on many years of clinical experience, that early treatment following a first episode of psychosis (FEP) provides the best chance for improving a patient’s subsequent outcome. In many cases, FEP is followed within 6 months to several years by a diagnosis of schizophrenia. While this is not the case in every instance, among those who do go on to manifest the clinical symptoms of schizophrenia, there is considerable variability in how each individual fares.

New research powered by machine learning now suggests a way of parsing the wide range of post-FEP trajectories, an approach which can help improve outcomes across the diversity of patients and perhaps also inform the development of personalized treatments. The team was led by 2023 BBRF Young Investigator David Benrimoh, M.D.C.M., FRCPC, of the Douglas Research Centre and McGill University, Canada. The team’s paper appeared in the journal npj Mental Health Research.

Schizophrenia is clinically characterized by two distinct types of clearly observable symptoms, called “positive” and “negative” symptoms. Positive symptoms include delusions, hallucinations, and peculiar or disorganized thoughts. Negative symptoms include alogia (difficulty generating coherent speech); flat affect; apathy; and lack of motivation or drive.

As noted by Dr. Benrimoh and his colleagues (who included two other BBRF grant recipients), “the trajectories and treatment responsiveness of positive and negative symptoms show stark differences.” Positive symptoms often respond to antipsychotic medications, but may worsen over time and can be a factor leading to hospitalizations when insufficiently treated. Negative symptoms generally don’t respond at all to widely prescribed first- and second-generation antipsychotics, but in some instances they can improve following FEP and can be targeted (with inconsistent results) by psychological and psychosocial treatments (which are often hard to access in the U.S.).

In the effort to develop personalized treatments targeting specific symptoms and their trajectories in FEP patients over time, the team set out to parse such trajectories “into coherent subgroups.” In the study just reported, they used a machine learning algorithm to identify three such subgroups. They then used a different machine learning technique to determine if the three subgroups could have been predicted in advance, when subjects analyzed in the study were first taken into a clinic after a first episode of psychosis.

Being able to predict illness trajectory at the time of initial patient intake “would enable clinicians to proactively allocate resources and modify treatment plans in a manner that is patient- or person-centered,” the team noted. “For example, if those at risk of treatment-resistant schizophrenia could be identified [in advance], more intensive interventions (such as intensive case management, psychosocial treatments, or early introduction of clozapine) could be considered in order to reduce the risk of the development of treatment resistance. In addition, the identification and characterization of symptom trajectory subgroups would provide important directions for future research aimed at developing subgroup-specific treatments or treatment protocols—a step towards personalized care.”

The patient data used by the team came from the Prevention and Early Intervention for Psychosis Program in Montreal, which serves individuals experiencing FEP. It encompassed the cases of 695 individuals aged 14 to 35 (mean age about 24), 70% of whom were male. 411 of these subjects were used in the effort to identify trajectory clusters.

When they entered the database, all subjects either had not yet received antipsychotic medicines or had received them for less than 1 month. They were tracked over 2 years, with data collected at nine timepoints—at “baseline” and at months 1, 2, 3, 6, 9, 12, 18, and 24. In addition to detailed demographic data, the team incorporated information about age at psychosis onset; length of the “prodrome,” or lead-up period to FEP, in which subthreshold symptoms may have been present; duration of untreated psychosis; timing of early symptoms; clinical diagnosis; adherence to medications; evaluations that measured social and occupational functioning; other psychiatric symptoms; and, importantly, scores from standard instruments for the assessment of positive and negative symptoms.

The primary analysis was based on a machine learning-based algorithm that measured the trajectories of scores on the positive- and negative-symptom assessments at all nine timepoints.

“We identified three clusters of psychotic symptom trajectories after a FEP,” the team reported. Each was “characterized by unique demographics, illness histories, longitudinal symptom patterns [i.e., over time], and prescribed antipsychotic doses.” The researchers found that they were also able, retrospectively, to “predict” which cluster each of the 411 study subjects would fall into, based solely on data that was available when they first entered the clinic for FEP.

They described each of the clusters as follows: relative to those in other clusters, patients in Cluster 1 (Low Symptoms, or “LS”) had lower positive and negative symptoms, lower antipsychotic medication dose, and relatively higher rate of affective disorder diagnoses, such as bipolar disorder, with psychotic features. Those in Cluster 2 (Low Positive, Persistent Negative, “LPPN”) had lower positive symptoms but persistent negative symptoms, and intermediate antipsychotic medication dose, relative to the others. Those in Cluster 3 had persistently high levels of both positive and negative symptoms (Persistent Positive and Negative Symptoms, “PPNS”) and higher antipsychotic doses.

Using data available at initial clinic intake alone, machine learning was able to predict which trajectory each patient was likely to follow with considerable accuracy. About 75% of the time, the model was able to determine, based solely on information gathered at "baseline," whether a given study subject fit within one of the three clusters describing illness trajectories.

Useful predictors differed according to cluster. Those in the LS cluster (Cluster 1) had lower levels of apathy, flat affect, and anhedonia/asociality compared with those in the LPPN Cluster (Cluster 2). Those in the PPNS cluster (Cluster 3) had more severe problems with hallucinations and disordered thought, and also were more likely to manifest hostile behavior.

Being able to predict symptom trajectory at clinical intake, the team said, “raises the possibility of secondary prevention approaches aimed at reducing treatment resistance.” It might mean that finding the best treatment approach would not have to wait for symptoms within each cluster to reach their peak; “rather, an approach could be developed with the aim of preventing progression along the predicted trajectory, by intervention earlier in the illness course—and, potentially, prior to illness onset.” A similar strategy, the researchers noted, is now being pursued in treating Alzheimer’s dementia—beginning treatments early and not waiting for full-blown illness to become manifest.

The results of this study open the way to a variety of follow-up research, the team suggested. This includes efforts to identify neurobiological differences distinguishing members of the three clusters, presuming that the current results are replicated by others. Brain scans, genetic testing, and other data might also be incorporated into the cluster profiles to better define them and potentially improve their utility.

The team also included: Martin Lepage, Ph.D., 2002 BBRF Young Investigator; and Ridha Joober, M.D., Ph.D., 2000 BBRF Young Investigator.