Background To prevent clozapine-induced agranulocytosis (CIA), patients’ white blood cell counts are closely monitored, with treatment stopped if the absolute neutrophil count (ANC) drops below 1.5×109/L. While effective, this approach has a high rate of false positives. This study aimed to develop a machine learning (ML) decision-making tool to better predict CIA risk using pattern-based criteria (two consecutive ANCs <0.5×109/L over ≥2 days).

Methods Using a ML technique [gradient-boosted decision trees (GDBT)] we analysed clinical data from 5,550 UK patients treated with clozapine: 2,190 controls with no history of neutropenia and 3,360 cases with at least one neutropenic event, including 358 with pattern-based CIA. Using haematological and demographic data from the current and three prior time windows, predictive models estimated the likelihood of CIA across four time-windows: 1 week, 2 weeks, 1 month, and 3 months respectively in advance. Model performance was evaluated using area under the receiver operator characteristic curve (AUROC), sensitivity, and specificity. We developed another model to predict baseline risk of CIA and compared performance with genetic tests. Explainability analyses identified key features influencing predictions.

Outcomes GDBT models demonstrated strong predictive performance: 1-week forecasting horizon: AUROC 0.99 [95% confidence interval (CI): 0.99–0.99]; 2 weeks: AUROC 0.97 [95% CI: 0.95–0.99]; 1 month: AUROC 0.91 [95% CI: 0.86–0.94]; 3 months: AUROC 0.90 [95% CI: 0.88–0.92]. The baseline model achieved better performance than current genetic tests, with high specificity and sensitivity at varying thresholds. Key discriminative features for CIA included age and baseline haematological values for longer forecasting horizons (1 and 3 months) and current haematological values and treatment duration for shorter horizons (1 and 2 weeks).

Interpretation ML models reliably predict CIA occurrence across short- and long-term horizons, potentially reducing the number of false positives with the current system. Implementation of ML models can reduce unnecessary treatment interruptions and the need for additional blood tests due to suspected agranulocytosis.

Funding The study did not receive direct funding.

Evidence before this study The only antipsychotic that is effective for treatmentresistant psychosis is clozapine. Tragically, many patients with treatment-resistant psychosis never receive clozapine treatment or receive it many years after “treatmentresistance”. A prominent reason for this is blood tests that are required to detect potential clozapine-induced agranulocytosis (CIA). Despite monitoring being effective, several patients have had to stop clozapine unnecessarily because of the current haematological criteria for discontinuation. In many of these patients, this has resulted in poor clinical and social outcomes. Additionally, many cases of agranulocytosis are identified late under the existing monitoring protocols. At present, there is no reliable way of predicting clozapine-induced agranulocytosis (CIA).

Added value of this study This is the first study to propose that a machine-learning decision tool can reliably predict CIA before it occurs in both the short term and long term.

Implications of all the available evidence Implementation of machine learning algorithms allow prediction of agranulocytosis so that clozapine can be appropriate stopped before it occurs. The algorithm can also prevent unnecessary stopping of clozapine and additional blood testing that is related to spurious blood results.

The authors have declared no competing interest.

The study did not receive direct funding.

I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained.

Yes

The details of the IRB/oversight body that provided approval or exemption for the research described are given below:

To accomplish our study aim, non-identifiable data were provided by the two largest manufacturers of clozapine and the associated haematological monitoring services in the UK; Clozaril (Mylan) monitored by Clozaril Patient Monitoring Service (CPMS) and Zaponex (Leyden Delta BV) monitored by Zaponex Treatment Access System (ZTAS). Data included haematological data such as ANC and WCC readings and patient demographic data such as age, ethnicity and gender. The CMPS dataset included national CNRD data on patients who recorded at least one neutropenic reading during clozapine treatment between 2000 to 2021. The ZTAS dataset included patients from one care setting in the UK (South London and Maudsley NHS Foundation Trust) who did or did not experi-ence neutropenia during clozapine treatment between 2004 to 2023. Data on comorbidities or medication us-age were not available. Ethical approval was not required according to the UK Health Research Authority (HRA).

I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals.

Yes

I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance).

Yes

I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable.

Yes

The data used in this study are not publicly available due to patient confidentiality but can be accessed upon reasonable request to the authors and subject to relevant ethical approvals.