Temporal patterns and risk factors of clozapine-associated neutropenia: a nationwide population-based cohort study

Summary

Background

Clozapine is the most effective antipsychotic for treatment-resistant schizophrenia, but neutropenia concerns continue to limit its use. Large real-world evidence on temporal risk patterns, re-initiation, and patient-level determinants remains limited.

Methods

We conducted a nationwide cohort study using linked claims and laboratory data from Taiwan's National Health Insurance database (2014–2023). New clozapine users and re-initiators after a ≥42 day interruption were identified. Neutropenia was classified as isolated minor, serious with treatment cessation, or serious without cessation, using absolute neutrophil count thresholds when available and white blood cell counts otherwise. Incidence rates were estimated per 1000 person-years. Temporal changes were examined using joinpoint regression, and risk factors using multivariable cause-specific Cox models with time-varying medication exposures.

Findings

Among 12,810 new users, isolated minor neutropenia occurred at 10.5 per 1000 person-years, and serious neutropenia leading to clozapine cessation at 1.7 per 1000 person-years. Joinpoint regression identified early inflection points at week 14 for isolated minor neutropenia, week 15 for serious neutropenia with cessation, and week 6 for serious neutropenia without cessation. Older age and lower baseline white blood cell count were consistently associated with higher risks. Concomitant medication associations were generally stronger for minor than serious events. Re-initiators showed risk patterns comparable to new users, and absolute neutrophil count-based sensitivity analyses yielded consistent results.

Interpretation

Clozapine-associated neutropenia was uncommon and concentrated early after initiation, with risk stratified by baseline haematological status and age. These findings support risk-informed monitoring strategies.

Funding

National Health Research Institutes, Taiwan.

Research in context.

Evidence before this study

We searched PubMed, EMBASE, and PsycINFO for studies published in English and Chinese from database inception to October 31, 2025, using the search terms “clozapine”, “neutropenia”, “agranulocytosis”, “incidence”, and “risk factors”. Previous research and meta-analyses have established that while clozapine is the most effective treatment for treatment-resistant schizophrenia, its use is hindered by the risk of neutropenia. However, most existing evidence is derived from Western populations, and large-scale real-world studies quantifying temporal risk trajectories and specific clinical predictors—especially in Asian populations—remain limited.

Added value of this study

This nationwide population-based cohort study of 12,810 new clozapine users provides one of the largest real-world evaluations of haematologic safety in an Asian population. Using linked health insurance and laboratory data, we identified a clear temporal risk trajectory: serious neutropenia leading to treatment cessation is uncommon and primarily occurs in the early phase of treatment, with a significant risk inflection at approximately week 15. Furthermore, we identified older age and low baseline white blood cell counts as the strongest clinical predictors for neutropenia across all severity levels.

Implications of all the available evidence

Our findings, combined with existing evidence, suggest that the risk of clozapine associated neutropenia is highly time-dependent and strongly stratified by individual baseline haematologic status and age. This supports a shift toward more personalized, risk-informed monitoring strategies rather than a one-size-fits-all approach. Intensified monitoring during the initial 15 weeks, particularly for high-risk subgroups, could optimize patient safety while potentially reducing the long-term monitoring burden and improving clozapine accessibility for patients with treatment resistant schizophrenia.

Introduction

Clozapine remains the most effective antipsychotic for treatment-resistant schizophrenia, with robust evidence demonstrating superior benefits in symptom reduction, relapse prevention, and mortality reduction.^1,2 Its distinctive pharmacological profile—including a low risk of extrapyramidal symptoms and potential protective effects against suicide and aggression—has also led to selective off-label use in conditions such as mood disorder, personality disorders, post-traumatic stress disorder, and psychosis associated with Parkinson's disease² Paradoxically, despite both its unique clinical advantages and long-standing guideline recommendations to initiate clozapine after inadequate response to two antipsychotic trials in schizophrenia,³ clozapine remains persistently underutilised in routine clinical practice worldwide.^4,5 Among the major barriers, haematologic toxicity—particularly neutropenia and agranulocytosis—has long been a central concern.

Cumulative prevalence estimates suggest that neutropenia occurs in roughly 3–4% and agranulocytosis in about 0.9% of clozapine-treated patients.⁶ Large registry and pharmacovigilance studies from US, UK, Europe, Australia and New Zealand further show that most serious neutropenic events leading to clozapine cessation occur in the first 18 weeks, with risk falling sharply thereafter and becoming very low after approximately two years of continuous treatment.7, 8, 9, 10 These safety concerns prompted the implementation of mandatory haematologic monitoring programs worldwide. However, international comparisons reveal substantial heterogeneity in monitoring frequency, absolute neutrophil count thresholds for treatment interruption, and policies regarding rechallenge.¹¹ Although such monitoring systems have been effective in reducing mortality, they also impose considerable logistical burdens, contributing to delayed initiation, premature discontinuation, and marked regional inequities in access to clozapine treatment.^5,12 Against this background, recent regulatory changes in the United States and Europe have moved toward less restrictive long-term clozapine blood monitoring, creating a need for population-based evidence on when neutropenia risk is concentrated and which patients remain at higher risk.13, 14, 15, 16

Despite this growing literature, several important gaps remain. First, attribution of neutropenia to clozapine versus background haematological variability or comorbid factors remains imperfect¹⁷; large-scale pharmacoepidemiologic studies are needed to better characterise risk patterns across diverse health systems. Second, optimal monitoring intensity and duration remain uncertain, particularly beyond the early high-risk period identified in contemporary registry studies.⁷ Third, evidence informing rechallenge after neutropenia remains limited,¹⁸ and large observational datasets may help to describe recurrence patterns in routine care.¹⁹

Using nationwide linked health-insurance claims and laboratory data, this study aimed to quantify the incidence, characterise the temporal pattern of neutropenia following clozapine initiation or re-initiation, and identify risk factors in routine clinical practice.

Methods

Data source

This population-based cohort study used data from Taiwan's National Health Insurance (NHI) claims database between 2014 and 2023. Established in 1995, the NHI program covers nearly 23 million individuals, representing approximately 99% of Taiwan's population. The database contains comprehensive information on demographic characteristics, medical diagnoses, and prescription records, with laboratory test results available upon special request.²⁰ Prescription records include both outpatient prescriptions and inpatient medication orders during hospital stays. In Taiwan's healthcare system, medications are typically dispensed directly through hospital-based or affiliated pharmacies without additional copayment at the point of dispensing. As a result, the vast majority of prescribed medications are expected to be dispensed, and prescription records are considered a close proxy for actual medication exposure in routine clinical practice.

Study participants

In Taiwan, clozapine is regulated under a haematologic monitoring program mandated by the Taiwan Food and Drug Administration. Blood counts are monitored weekly for the first 18 weeks after initiation and at least every four weeks thereafter. The guideline allows monitoring using either total white blood cell count (WBC) or absolute neutrophil count (ANC); ANC measurement is not mandatory, and treatment decisions may be based on WBC alone. Clozapine must be discontinued if WBC falls below 2500/μL or ANC falls below 1000/μL.

We identified the study cohort from National Health Insurance (NHI) claims records, comprising 46,594,966 claims from 1,613,751 individuals diagnosed with major psychiatric disorders between 2014 and 2023. After restricting the sample to individuals with complete sociodemographic information, 1,564,922 patients remained. Among these individuals, 41,735 had at least one recorded clozapine prescription. After exclusion of 11,546 laboratory records with invalid values (extreme measurements or missing units), 327,968 valid WBC or ANC measurements from clozapine-treated patients were available for analysis.

We defined new clozapine users as those initiating treatment after 2015, with the first prescription following a 365-day washout period without prior clozapine exposure, yielding 22,136 individuals. Restricting the cohort to patients with an available baseline WBC or ANC measurement and at least one follow-up WBC test resulted in a final analytic cohort of 12,810 new clozapine users (Fig. 1).

Fig. 1.

Flow diagram of study cohort selection and eligibility criteria.

Clozapine treatment episodes were identified using prescription records from both outpatient and inpatient settings. Discontinuation was defined as a gap of more than 42 days without a clozapine prescription. Each treatment episode ended on the date of the last prescription plus its days of supply. Treatment duration was calculated from the first prescription date to the end date of the final prescription within each continuous episode.

For the re-initiator cohort, we included patients restarting after a ≥42 day discontinuation. Most experts recommend that if clozapine treatment has been interrupted for 4 weeks or longer, weekly WBC or ANC monitoring should be resumed for the next 18 weeks. However, because a gap in prescription does not necessarily indicate complete discontinuation, we defined an interruption of 42 days as the threshold for identifying re-initiators. Individuals could contribute more than one clozapine treatment episode over the study period. Accordingly, a given patient could enter the analysis first as a new user and later contribute one or more re-initiation episodes after treatment interruption.

Outcome measures

Neutropenia events were classified into three mutually exclusive categories using ANC thresholds; when ANC was unavailable, WBC counts were used as a proxy. Isolated minor neutropenia was defined as ANC 1000–1500/μL or WBC 2500–3000/μL. Serious neutropenia was defined as ANC <1000/μL or WBC <2500/μL and further classified as with or without clozapine cessation within 6 weeks. Events without cessation were typically attributed to alternative causes. The WBC cutoffs were consistent with Taiwanese clozapine monitoring guidelines. In preliminary analyses, among patients with concurrent ANC and WBC measurements and ANC <1000/μL, the neutrophil fraction (ANC/WBC) had a mean (SD) of 42.7% (24.9%), providing empirical support for the use of these WBC thresholds as a proxy.

Covariates

Age at clozapine initiation was categorised as <35, 35–44, 45–54, or ≥55 years, as preliminary analyses suggested a non-linear relationship between age and the outcome. Sex was recorded in the NHI claims database as male or female. Gender identity was not available. Race and ethnicity were not available in the NHI claims database and were therefore not analysed. Socioeconomic status was proxied by monthly insurance salary (≤11,100 [minimum insured salary], 11,101–24,000, or >24,000 New Taiwan Dollars). The residential area was classified as urban, suburban, or rural according to the degree of urbanisation.

Psychiatric conditions at the index date were classified as schizophrenia, bipolar disorder, depressive disorders, or other diagnoses using claims data from the National Health Insurance database based on ICD-9-CM and ICD-10-CM codes. Schizophrenia was defined by ICD-9-CM 295.xx and ICD-10-CM F20.x and F25.x. Bipolar disorder was identified using ICD-9-CM codes 296.0, 296.1, and 296.4–296.8, and ICD-10-CM codes F30 and F31. Depressive disorders were defined by ICD-9-CM 296.2 and 296.3, and ICD-10-CM F32 and F33. Other diagnoses included all psychiatric disorders coded as ICD-9-CM 290–319 or ICD-10-CM F codes that were not classified as schizophrenia, bipolar disorder, or depressive disorders. Diagnoses required at least one inpatient or two outpatient claims. Comorbidity burden was assessed using the Charlson Comorbidity Index (CCI) during the year preceding clozapine initiation and categorised as 0, 1, or ≥2²¹ Baseline WBC count before clozapine initiation was grouped as <4000, 4000–5999, 6000–7999, or ≥8000 cells/μL.

Concomitant medications were assessed both as baseline characteristics and as time-varying exposures. Medication use in the year prior to clozapine initiation was used to characterise baseline patient profiles, whereas in Cox regression models, medications were treated as time-varying exposures using 28 day rolling windows based on prescriptions in the preceding month to better align exposure with the risk period and capture changes during follow-up. Psychotropic medications included benzodiazepines, antidepressants, and non-clozapine antipsychotics prescribed in the year prior to initiation. Other commonly prescribed medications during this period included angiotensin-converting enzyme inhibitors or angiotensin receptor blockers (ACEI/ARB), beta-blockers, calcium channel blockers, diuretics, lipid-lowering agents, antidiabetic medications, nonsteroidal anti-inflammatory drugs (NSAIDs), corticosteroids, carbamazepine, valproic acid, and other antiepileptic drugs. Among re-initiators, patients were additionally classified according to a documented history of neutropenia. All medications were identified using Anatomical Therapeutic Chemical (ATC) classification codes from the National Health Insurance prescription database (see Supplementary Table S1 for full code lists).

Statistical analysis

We conducted survival analyses to evaluate the risk of neutropenia among clozapine-treated patients. Follow-up began at the date of clozapine initiation (or re-initiation) and ended at the earliest of a neutropenic event, discontinuation of clozapine for non-neutropenic causes, or the last available WBC test at the census date. For the re-initiation cohort, multiple treatment episodes from the same individual were allowed. Each eligible re-initiation episode was treated as a separate risk interval, and robust sandwich variance estimators clustered at the individual level were used to account for within-person correlation.

Incidence estimation

Crude incidence rates were calculated for isolated minor neutropenia, serious neutropenia leading to cessation, and serious neutropenia without cessation. Rates were derived by dividing the number of events by the corresponding person-years of follow-up and expressed per 1000 person-years. Incidence was further stratified by demographic, clinical, and treatment characteristics, including age, sex, psychiatric indication, baseline WBC, and concomitant medication use.

Joinpoint regression

Joinpoint regression was used to identify inflection points in the weekly incidence of neutropenia following clozapine initiation. Models assumed a Poisson distribution for event counts, with the log of person-time included as an offset to estimate incidence rates. This approach is standard for analysing temporal trends in aggregated incidence data. For each outcome and cohort, models with 0, 1, or 2 joinpoints were fitted to capture potential changes in incidence over time. To limit overfitting, candidate joinpoints were required to be separated by a minimum number of weeks, and a fast grid-search algorithm was applied for two-joinpoint models. Model selection was based on the Bayesian Information Criterion (BIC), with lower values indicating better fit; when BIC differences were small, the more parsimonious model was favoured. Final models were chosen based on statistical fit, consistency across cohorts, and clinical interpretability in relation to established clozapine monitoring guidelines.

Risk factor analysis

Cox proportional hazards models were used to estimate adjusted hazard ratios (HRs) and 95% confidence intervals (CIs) for any neutropenia event. Baseline demographic and clinical covariates were defined at clozapine initiation. Concomitant medications in the Cox models were treated as time-varying exposures using 28-day rolling windows based on prescriptions in the preceding month. In analyses of isolated minor neutropenia, serious neutropenia with cessation, and serious neutropenia without cessation, each outcome was treated as a competing event for the others. Death was also considered a competing risk. We therefore used cause-specific hazard models, censoring individuals at death or at the occurrence of a competing neutropenia subtype, to estimate the instantaneous risk of the outcome of interest among those who remained event-free.

Sensitivity analysis

Given that ANC measurement is not mandatory, we used WBC measures as a proxy. Among tests with concurrent ANC and WBC measurements, WBC-based thresholds showed high overall agreement with ANC-defined neutropenia (98.9%), with a Cohen's κ of 0.61 (Supplementary Table S2). When evaluated against ANC as the reference standard, WBC thresholds demonstrated high specificity (99.6%) and negative predictive value (99.2%), but lower sensitivity (54.1%), indicating that a proportion of ANC-defined neutropenia events may not be captured using WBC alone. The positive predictive value was 71.5%.

Therefore, we conducted a sensitivity analysis restricted to individuals with available baseline and at least one follow-up ANC measurement. Neutropenia was redefined using standard ANC thresholds (minor: ANC <1500/μL; serious: ANC <1000/μL), and all analyses were repeated using these ANC-based definitions. Restricting the cohort to individuals with ANC data reduced the sample to 10,510 new users.

All analyses were conducted using SAS version 9.4 (SAS Institute Inc., Cary, NC, USA). Statistical significance was defined as a two-sided p value < 0.05.

Statement of ethics

This study was approved by the Research Ethics Committee of the National Health Research Institutes, Taiwan (EC1140111-E; approved on March 5, 2025). The requirement for informed consent was waived because the study used encrypted, de-identified administrative claims and laboratory data.

Role of the funding source

The funders had no role in study design, data collection, data analysis, interpretation, or writing of the report.

Results

Baseline characteristics

Among 22,136 new clozapine users, 12,810 met the inclusion criteria, defined by having at least two WBC tests. The remaining 9326 patients (42.1%) were excluded due to missing baseline WBC/ANC data (Supplementary Table S3). Compared with included patients, those without laboratory data were similar in age, urbanisation, and comorbidity burden, but differed in clinical profile and treatment patterns. Included patients were more likely to have schizophrenia and to receive more intensive antipsychotic and psychotropic treatment (e.g., benzodiazepines, valproic acid), whereas excluded patients had a higher prevalence of depressive disorders and antidepressant use, suggesting that the analytic cohort represents a more closely monitored and clinically severe subgroup.

Among included new users, 24.5% were aged <35 years, 23.2% were 35–44 years, 22.8% were 45–54 years, and 29.4% were ≥55 years; 52.5% were female. Schizophrenia was the predominant indication (67.8%), followed by bipolar disorder (13.2%) and depressive disorders (11.3%). Most patients had no comorbidities in the prior year (CCI = 0, 93.0%). Baseline WBC counts were ≥ 8000/μL in 38.3%, 6000–7999/μL in 33.2%, 4000–5999/μL in 24.9%, and <4000/μL in 3.7% (Table 1).

Table 1.

Baseline characteristics of clozapine treatment episodes among new users and re-initiators.

	New users only (n = 12,810)	Re-initiators (N = 6153)	SMD
Age, years
<35	3143 (24.5)	1025 (16.7)	0.19
35–44	2976 (23.2)	1598 (26.0)	0.07
45–54	2923 (22.8)	1813 (29.5)	0.15
≥55	3768 (29.4)	1717 (27.9)	0.03
Sex
Male	6089 (47.5)	3124 (50.8)	0.07
Female	6721 (52.5)	3029 (49.2)	0.07
Income, proxy by insurance salary (NTD)
≤11,100	7253 (56.6)	3880 (63.1)	0.13
11,101–24,000	2620 (20.5)	1133 (18.4)	0.05
>24,000	2937 (22.9)	1140 (18.5)	0.11
Urbanisation
Urban	6467 (50.5)	3104 (50.5)	0.00
Suburban	4945 (38.6)	2389 (38.8)	0.00
Rural	1398 (10.9)	660 (10.7)	0.01
Psychiatric indications, at index date
Schizophrenia	8687 (67.8)	4821 (78.4)	0.24
Bipolar disorders	1692 (13.2)	647 (10.5)	0.08
Depressive disorders	1446 (11.3)	349 (5.7)	0.20
Others	985 (7.7)	336 (5.5)	0.09
Previous Clozapine Exposure
No history of neutropenia		6107 (99.3)
With a history of neutropenia		46 (0.8)
Baseline WBC
<4000	470 (3.7)	218 (3.5)	0.01
4000–5999	3183 (24.9)	1609 (26.2)	0.03
6000–7999	4257 (33.2)	1998 (32.5)	0.01
≥8000	4900 (38.3)	2328 (37.8)	0.01
CCI, in the past year
0	11,916 (93.0)	5843 (95.0)	0.08
1	530 (4.1)	184 (3.0)	0.06
≥2	364 (2.8)	126 (2.1)	0.05
Medications, in the past year
Psychotropic Medications
Benzodiazepines	11,559 (90.2)	5145 (83.6)	0.20
Antidepressants	7067 (55.2)	3011 (48.9)	0.13
Antipsychotics
≤1	1587 (12.4)	1470 (23.9)	0.30
2–3	5046 (39.4)	1748 (28.4)	0.23
4–5	4146 (32.4)	1739 (28.3)	0.09
≥6	2031 (15.9)	1196 (19.4)	0.09
ACEI/ARB	1990 (15.5)	801 (13.0)	0.07
Beta-blockers	6402 (50.0)	2984 (48.5)	0.03
Calcium channel blockers	2523 (19.7)	1182 (19.2)	0.01
Diuretics	1242 (9.7)	704 (11.4)	0.06
Lipid-lowering agents	1893 (14.8)	886 (14.4)	0.01
Antidiabetic medications	1914 (14.9)	1076 (17.5)	0.07
NSAIDs	8082 (63.1)	3650 (59.3)	0.08
Corticosteroids	3168 (24.7)	1529 (24.9)	0.00
Carbamazepine	423 (3.3)	225 (3.7)	0.02
Valproic acid	5030 (39.3)	2594 (42.2)	0.06
Other antiepileptics	1372 (10.7)	753 (12.2)	0.05

Baseline medication use was assessed during the year before clozapine initiation or re-initiation.

Values are presented as a number (%).

WBC denotes white blood cell count; SMD, standardised mean difference; NTD, New Taiwan dollars; ACE, angiotensin-converting enzyme; ARB, angiotensin receptor blocker.

Baseline medication use was assessed during the year before clozapine initiation. An SMD greater than 0.10 was considered to indicate a meaningful imbalance.

A total of 6153 re-initiation treatment episodes from 4850 individuals met eligibility criteria. Most individuals contributed only one re-initiation episode (n = 3,918, 80.8%); 680 (14.0%) contributed two episodes, 181 (3.7%) contributed three episodes, 43 (0.9%) contributed four episodes, and 28 (0.6%) contributed five or more episodes. Therefore, repeated episodes were allowed, but most re-initiation observations came from individuals with a single re-initiation episode. Among re-initiators, the duration from first clozapine initiation to first discontinuation had a mean of 781.2 days (SD, 658.7) and a median of 582 days (IQR, 856). The interval from discontinuation to re-initiation had a mean of 741.4 days (SD, 649.2) and a median of 519.5 days (IQR, 795), indicating a right-skewed distribution in which most patients restarted within approximately 1.5 years, while some experienced substantially longer delays. Most re-initiators (99.3%) had no documented history of neutropenia prior to re-initiation. Their baseline characteristics were broadly similar to new users, although they were older, had lower income, were more likely to have schizophrenia, and were less likely to use benzodiazepines or antidepressants.

Incidence of neutropenia

During follow-up, there were 524 isolated minor neutropenia events (10.5 per 1000 person-years), 84 serious events with clozapine cessation (1.7 per 1000 person-years), and 144 serious events without cessation (2.9 per 1000 person-years) (Table 2). Among re-initiators, 268 isolated minor events occurred (12.5 per 1000 person-years), along with 48 serious events with cessation (2.2 per 1000 person-years) and 72 serious events without cessation (3.4 per 1000 person-years) (Supplementary Table S4).

Table 2.

Incidence rates and adjusted hazard ratios for clozapine-associated neutropenia among new users.

	Person-Years	Isolated minor neutropenia			Serious neutropenia leading to cessation			Serious neutropenia without cessation
		Event number	Incidence	Adjusted HR (95% CI)	Event number	incidence	Adjusted HR (95% CI)	Event number	incidence	Adjusted HR (95% CI)
Overall	50,062	524	10.5		84	1.7		144	2.9
Age, years
<35	13,181	68	5.1	Referent	6	0.4	Referent	18	1.3	Referent
35–44	12,678	98	7.6	1.47 (1.06, 2.04)	18	1.4	3.12 (1.23, 7.91)	39	3.0	2.24 (1.26, 3.98)
45–54	11,868	139	11.6	1.94 (1.42, 2.64)	17	1.4	2.69 (1.04, 6.92)	41	3.4	2.27 (1.29, 4.01)
≥55	12,335	219	17.5	3.11 (2.31, 4.20)	43	3.3	6.41 (2.59, 15.90)	46	3.6	2.76 (1.55, 4.93)
Sex
Male	24,176	235	9.6	0.85 (0.70, 1.03)	44	1.8	1.35 (0.86, 2.14)	74	3.0	1.09 (0.76, 1.55)
Female	25, 886	289	11.1	Referent	40	1.5	Referent	70	2.6	Referent
Income, proxy by insurance salary (NTD)
≤11,100	28,465	351	12.2	Referent	50	1.7	Referent	89	3.0	Referent
11,101–24,000	12,794	88	6.8	0.64 (0.50, 0.82)	14	1.1	0.57 (0.31, 1.07)	37	2.8	0.99 (0.67, 1.48)
>24,000	8803	85	9.6	0.80 (0.61, 1.05)	20	2.2	1.07 (0.61, 1.88)	18	2.0	0.79 (0.46, 1.37)
Urbanisation
Urban	24,997	267	10.6	Referent	40	1.6	Referent	65	2.5	Referent
Suburban	19,626	208	10.5	1.12 (0.92, 1.36)	31	1.5	1.14 (0.70, 1.85)	58	2.9	1.16 (0.81, 1.67)
Rural	5439	49	8.9	0.97 (0.70, 1.35)	13	2.3	1.77 (0.92, 3.41)	21	3.8	1.47 (0.89, 2.44)
Psychiatric indications, at index date
Schizophrenia	35,300	365	10.2	Referent	44	1.2	Referent	105	2.9	Referent
Bipolar disorders	6340	49	7.7	0.76 (0.56, 1.04)	16	2.5	1.82 (0.96, 3.45)	14	2.2	0.72 (0.40, 1.30)
Depressive disorders	4949	45	9.0	0.93 (0.65, 1.32)	17	3.4	2.41 (1.20, 4.85)	8	1.6	0.52 (0.24, 1.16)
Others	3474	65	18.4	1.42 (1.05, 1.92)	7	1.9	1.22 (0.53, 2.83)	17	4.7	1.36 (0.78, 2.38)
Baseline WBC
<4000	1945	120	61.7	13.29 (9.63, 18.36)	19	9.8	11.42 (5.54, 23.53)	35	18.0	9.90 (5.78, 16.97)
4000–5999	11,641	213	18.1	4.04 (3.09, 5.29)	27	2.2	2.75 (1.47, 5.14)	38	3.1	1.89 (1.17, 3.07)
6000–7999	16,844	114	6.7	1.63 (1.22, 2.19)	22	1.3	1.56 (0.82, 2.96)	40	2.3	1.45 (0.90, 2.34)
≥8000	19,632	77	3.9	Referent	16	0.8	Referent	31	1.6	Referent
CCI, in the past year
0	44,230	498	11.1	Referent	75	1.6	Referent	132	2.9	Referent
1	3617	14	3.8	0.46 (0.26, 0.79)	4	1.1	0.42 (0.15, 1.17)	7	1.9	0.33 (0.15, 0.73)
≥2	2214	12	5.3	0.63 (0.34, 1.16)	5	2.2	0.56 (0.21, 1.47)	5	2.2	0.32 (0.12, 0.83)
Time-Varying Medication Use, in the past month
Benzodiazepines	23,589	330	14.0	1.68 (1.39, 2.02)	72	3.1	1.28 (0.79, 2.05)	108	4.6	1.16 (0.79, 1.70)
Antidepressant	17,343	174	10.0	0.94 (0.76, 1.15)	47	2.7	0.85 (0.50, 1.43)	60	3.5	1.10 (0.72, 1.68)
Non-clozapine antipsychotics	28,391	372	13.1	1.49 (1.21, 1.85)	81	2.9	1.67 (1.00, 2.80)	138	4.9	1.02 (0.70, 1.48)
ACEI/ARB	3750	53	14.1	1.06 (0.74, 1.53)	17	4.5	1.37 (0.69, 2.73)	16	4.3	0.93 (0.44, 1.96)
Beta-blockers	16,388	179	10.9	0.90 (0.73, 1.11)	41	2.5	1.06 (0.63, 1.78)	59	3.6	0.95 (0.64, 1.40)
Calcium channel blockers	4714	76	16.1	0.81 (0.58, 1.14)	25	5.3	1.00 (0.52, 1.94)	24	5.1	0.79 (0.42, 1.48)
Diuretics	2234	48	21.5	1.80 (1.19, 2.71)	19	8.5	1.34 (0.51, 3.52)	19	8.5	2.24 (1.04, 4.85)
Lipid-lowering agents	4738	40	8.4	0.76 (0.51, 1.13)	12	2.5	0.83 (0.33, 2.12)	14	3.0	0.67 (0.30, 1.48)
Antidiabetic medications	5907	56	9.5	1.19 (0.86, 1.65)	17	2.9	1.06 (0.49, 2.29)	25	4.2	1.59 (0.90, 2.82)
NSAID	15,375	174	11.3	0.92 (0.70, 1.21)	53	3.4	1.85 (1.06, 3.21)	57	3.7	0.87 (0.51, 1.47)
Corticosteroids	3396	91	26.8	1.77 (1.22, 2.58)	28	8.2	3.18 (1.69, 5.99)	26	7.7	2.33 (1.21, 4.46)
Carbamazepine	870	13	14.9	1.55 (0.83, 2.88)	6	6.9	3.32 (1.16, 9.52)	7	8.0	2.77 (1.11, 6.91)
Valproic acid	12,177	173	14.2	2.06 (1.68, 2.51)	35	2.9	1.81 (1.09, 2.99)	57	4.7	1.43 (0.96, 2.14)
Other antiepileptics	2548	44	17.3	1.13 (0.73, 1.76)	11	4.3	1.05 (0.39, 2.85)	13	5.1	1.44 (0.65, 3.21)

Abbreviations: HR, hazard ratio; CI, confidence interval; WBC, white blood cell count; CCI, Charlson Comorbidity Index; ACEI/ARB, angiotensin-converting enzyme inhibitor or angiotensin receptor blocker; NSAID, nonsteroidal anti-inflammatory drug.

Incidence rates are expressed per 1000 person-years.

Bold text indicates statistical significance.

Adjusted hazard ratios were estimated using multivariable cause-specific Cox proportional hazards models. Models were adjusted for age, sex, insurance salary, urbanisation level, psychiatric diagnosis, baseline WBC category, comorbidity burden, and all medication variables listed in the table.

Medication exposures were modelled as time-varying covariates using 28-day rolling prescription windows. For these exposures, person-years represent exposed person-time; hazard ratios compare exposed with unexposed person-time and were mutually adjusted for the other listed medications and baseline covariates.

Joinpoint regression

Joinpoint regression identified a consistent early inflection in the weekly incidence of all neutropenia outcomes following clozapine initiation. As shown in Table 3, a one-joinpoint model provided the best fit for all outcomes among new users. The estimated joinpoints occurred at week 14 for isolated minor neutropenia, week 15 for serious neutropenia with cessation, and week 6 for serious neutropenia without cessation (Fig. 2a–c).

Table 3.

Comparison of joinpoint regression models for weekly incidence of clozapine-associated neutropenia among new users and re-initiators.

Neutropenia type	Model	New users Join1	New users Join1	New users Join1	Re-initiators Join1	Re-initiators Join1	Re-initiators Join1
Isolated minor neutropenic events	0-join	–	–	−904.1	–	–	−835.16
	1-join	14	–	−1067.9	11	–	−874.25
	2-join	11	46	−1069.9	11	126	−859.32
Serious neutropenic events leading to cessation	0-join	–	–	−1327.5	–	–	−1064.2
	1-join	15	–	−1368.5	146	–	−1064.6
	2-join	16	51	−1367.5	131	141	−1057.1
Serious neutropenic events without cessation	0-join	–	–	−1302.0	–	–	−1082.7
	1-join	6	–	−1380.0	19	–	−1086.9
	2-join	11	31	−1361.9	21	71	−1072.5

Bolded values indicate the model retained for interpretation, considering both BIC and parsimony.

Although some joinpoint models yielded marginally lower BIC values for selected outcomes, the improvements were minimal (ΔBIC ≤2) and lacked clinical relevance; therefore, zero- or one-joinpoint models were retained for parsimony and interpretability.

Fig. 2.

a. Weekly incidence of clozapine-associated neutropenia among new users identified by joinpoint regression. b. Weekly incidence of serious neutropenic events leading to cessation among new users identified by joinpoint regression. c. Weekly incidence of serious neutropenic events without cessation among new users identified by joinpoint regression.

Among re-initiators, a one-joinpoint model identified an early inflection at week 11 for isolated minor neutropenia. For serious neutropenia with cessation, which was a rare event, the zero-joinpoint model was retained for parsimony, as the one-joinpoint model yielded only a marginal improvement in BIC (ΔBIC = 0.4) and placed the joinpoint at week 146. For serious neutropenia without cessation, a one-joinpoint model identified an inflection at week 19 (Table 3).

Overall, despite some variation in joinpoint timing between cohorts, the pattern was consistent, with the primary increase in clozapine-associated neutropenia risk occurring during the early treatment phase.

Risk factor analysis

Among new clozapine users, older age was consistently associated with higher risks of neutropenia. Compared with patients aged <35 years, those aged ≥55 years had higher risks of isolated minor neutropenia (adjusted hazard ratio [aHR] 3.11, 95% CI 2.31–4.20), serious neutropenia leading to treatment cessation (aHR 6.41, 95% CI 2.59–15.90), and serious neutropenia without cessation (aHR 2.76, 95% CI 1.55–4.93). The incidence of serious neutropenia leading to treatment cessation was highest in this age group, at 3.3 per 1000 person-years.

Baseline haematologic status was the strongest predictor: compared with baseline WBC ≥8000 cells/μL, baseline WBC <4000 cells/μL was associated with markedly higher risks of isolated minor neutropenia (aHR 13.29, 95% CI 9.63–18.36), serious neutropenia leading to cessation (aHR 11.42, 95% CI 5.54–23.53), and serious neutropenia without cessation (aHR 9.90, 95% CI 5.78–16.97).

Psychiatric indication and concomitant medications were also associated with neutropenia risk. Compared with schizophrenia, other psychiatric diagnoses were associated with isolated minor neutropenia (aHR 1.42, 95% CI 1.05–1.92), while depressive disorders were associated with serious neutropenia leading to cessation (aHR 2.41, 95% CI 1.20–4.85). In time-varying medication analyses, valproic acid was associated with isolated minor neutropenia (aHR 2.06, 95% CI 1.68–2.51) and serious neutropenia leading to cessation (aHR 1.81, 95% CI 1.09–2.99). Carbamazepine was associated with both serious neutropenia leading to treatment cessation (aHR 3.32, 95% CI 1.16–9.52) and serious neutropenia without cessation clozapine (aHR 2.77, 95% CI 1.11–6.91). Benzodiazepines and non-clozapine antipsychotics were associated with increased risks of isolated minor neutropenia (aHR 1.68, 95% CI 1.39–2.02; and aHR 1.49, 95% CI 1.21–1.85, respectively). Diuretics were associated with isolated minor neutropenia (aHR 1.80, 95% CI 1.19–2.71) and serious neutropenia without cessation (aHR 2.24, 95% CI 1.04–4.85), while corticosteroids were associated with all three outcomes (Table 2).

Among re-initiators, incidence patterns and risk–factor associations were broadly similar to those observed among new users (Supplementary Table S4). Older age and lower baseline WBC count remained the most consistent predictors of neutropenia. Valproic acid use was associated with an increased risk of isolated minor neutropenia, whereas corticosteroid and diuretic use were associated with serious neutropenia leading to treatment cessation. Patients with a prior history of neutropenia had numerically higher rates of recurrent neutropenia, particularly of serious neutropenia without cessation, but these estimates were based on very small numbers of cases and should be interpreted cautiously.

Sensitivity analysis

Sensitivity analyses using ANC-defined outcomes yielded patterns broadly consistent with the primary WBC-based analysis (Supplementary Table S5). Although absolute incidence rates were lower when ANC thresholds were used, the main risk–factor associations remained stable. Older age, lower baseline WBC count, and valproic acid use continued to be associated with higher risks of neutropenia. These findings suggest that the main conclusions were robust to potential outcome misclassification from using WBC-based definitions.

Discussion

In this nationwide cohort study of patients treated with clozapine, we characterised the incidence, temporal pattern of neutropenia, and risk factors in routine clinical practice using linked claims and laboratory data. Three principal findings emerged. First, neutropenia was uncommon overall, with serious events requiring clozapine cessation rarely occurring. Second, joinpoint regression revealed a consistent early inflection in neutropenia risk—occurring approximately 14–15 weeks after clozapine initiation—after which incidence declined substantially and remained low. Third, risk of all categories of neutropenia was strongly stratified by patient characteristics, particularly older age and low baseline WBC count. Several time-varying concomitant medications were also associated with neutropenia risk: benzodiazepines, non-clozapine antipsychotics, diuretics, corticosteroids, valproic acid, and carbamazepine showed outcome-specific associations. These patterns were broadly consistent among re-initiators and in sensitivity analyses restricted to ANC-defined outcomes.

The incidence estimates observed in this study are broadly consistent with those reported in large registry and pharmacovigilance studies from Europe, North and South America, and Australasia, which have shown that clozapine-associated neutropenia is rare and concentrated early in treatment.6, 7, 8, 9, 10^,22 This study provides large-scale evidence from Taiwan, a non-Western health-care system with a predominantly East Asian population, suggesting that the reduction in risk of neutropenia, both minor and serious leading to cessation, is independent of ethnicity.

Importantly, re-initiators did not materially increase the overall incidence of neutropenia. However, these findings should be interpreted with caution. Individuals with a prior history of neutropenia represented a small subgroup, and although recurrence rates were numerically higher, the estimates were imprecise due to limited event numbers. Accordingly, while our findings provide some reassurance regarding re-initiation at the population level, they do not establish the safety of rechallenge among individuals with prior neutropenia. Further studies with larger samples or pooled data are needed to better characterise recurrence risk and inform clinical decision-making in this high-risk subgroup.

A key contribution of this study lies in its explicit modelling of the temporal dynamics of neutropenia risk using joinpoint regression. Across neutropenia categories, we identified a consistent early inflection in incidence among new users, with broadly similar early patterns observed among re-initiators. These findings indicate that clozapine-associated neutropenia risk is concentrated in the early treatment phase, with incidence declining thereafter. For serious neutropenia leading to treatment cessation, this inflection occurred at week 15, closely aligning with the early intensive monitoring period recommended in many clinical guidelines.^23,24 Unlike prior studies that relied on descriptive incidence curves or arbitrary time intervals,^8,9 our joinpoint approach offers a data-driven method to identify when risk trajectories change, thereby strengthening the evidence base underlying early monitoring recommendations. At the same time, the persistence of low but non-zero risk beyond the early phase underscores the need for continued, albeit potentially less intensive, surveillance.

An important consideration is differential surveillance. In Taiwan, haematologic monitoring is weekly during the first 18 weeks of clozapine treatment and less frequent thereafter, which may increase the early detection of transient or mild abnormalities and inflate incidence estimates after initiation. However, the early inflection in neutropenia risk is unlikely to be solely due to surveillance. The timing aligns with prior registry and pharmacovigilance studies across settings with different monitoring practices. Therefore, the findings likely reflect both biological risk and detection effects.

These findings should be interpreted in the context of recent regulatory changes in clozapine blood monitoring. In the United States, the FDA eliminated the Clozapine Risk Evaluation and Mitigation Strategy (REMS) program to reduce barriers to clozapine access. In Europe, clozapine monitoring requirements have also been formally revised under the European Medicines Agency regulatory framework. The revised recommendations retain intensive early monitoring, with weekly blood monitoring during the first 18 weeks and monthly monitoring until completion of the first year, but allow monitoring to be reduced to every 12 weeks after one year and annually after two years in patients without a history of neutropenia. These changes are consistent with accumulating evidence that the risk of clozapine-associated severe neutropenia is concentrated early in treatment and becomes substantially lower during long-term therapy.7, 8, 9, 10 Our findings add population-based evidence from an East Asian health-care system, showing a similar early concentration of risk and strong stratification by baseline haematological status and age.13, 14, 15, 16

However, our results should be interpreted as descriptive evidence rather than direct guidance for changing monitoring schedules. This study did not evaluate alternative monitoring protocols or assess the clinical consequences of de-escalated monitoring. Nevertheless, by demonstrating that neutropenia risk is time-dependent and strongly stratified by baseline risk factors, our findings provide empirical support for future studies evaluating risk-stratified or de-escalated monitoring strategies in carefully selected populations.

Age emerged as a consistent determinant of neutropenia risk, with patients aged 55 years and older experiencing the highest incidence of serious neutropenia leading to treatment cessation. This finding is consistent with prior reports and may reflect age-related differences in haematopoietic reserve,²⁵ comorbidity burden, or cumulative medication exposure.²⁶ Baseline haematological status was the strongest predictor of subsequent neutropenia, with markedly elevated risks observed among patients with low pre-treatment WBC counts, consistent with previous study.^7,17 This gradient was evident across all neutropenia subtypes and persisted in sensitivity analyses restricted to ANC-defined outcomes, underscoring the clinical importance of baseline laboratory assessment for risk stratification.

Lower socioeconomic status was associated with a higher risk of minor neutropenia only. This finding likely reflects greater comorbidity burden, poorer baseline health and nutrition, which may reduce haematopoietic reserve.²⁷ Socioeconomic disadvantage may also impede timely access to care and adherence to monitoring.²⁸

Several time-varying concomitant medications were associated with neutropenia risk. The association with valproic acid is biologically plausible and consistent with prior concerns regarding additive haematological toxicity during clozapine treatment.²⁹ Carbamazepine was associated with serious neutropenia, which is also clinically plausible given its known potential to cause neutropenia and rare agranulocytosis; concomitant use with clozapine therefore warrants particular caution.³⁰ Non-clozapine antipsychotics and benzodiazepines were associated mainly with isolated minor neutropenia, but these findings may reflect psychiatric severity, polypharmacy, or residual confounding rather than direct myelotoxicity. Diuretics and corticosteroids were also associated with selected serious outcomes; however, these drugs may act as markers of acute medical illness, inflammatory disease, or increased laboratory surveillance. Overall, these medication associations should be interpreted as clinically useful risk markers rather than definitive causal effects, especially because concomitant treatment was time-varying and may reflect changing clinical status during follow-up.

The analysis of re-initiators merits particular consideration. Patients restarting clozapine after treatment interruption showed incidence patterns similar to those of new users, suggesting that re-initiation per se does not substantially alter risk trajectories. However, individuals with a documented history of neutropenia exhibited higher recurrence rates, although the number of such cases was small and estimates were imprecise. Larger datasets or pooled analyses will be required to better characterise recurrence patterns and inform rechallenge decisions.

Limitations

Several limitations should be acknowledged. First, this study did not include all clozapine users in Taiwan. Most excluded individuals lacked at least two recorded WBC or ANC measurements. Compared with included patients, excluded patients had a higher prevalence of depressive disorders and other non-schizophrenia diagnoses, as well as greater antidepressant use, suggesting that off-label or non-standard clozapine use may be less consistently captured in the laboratory database. Conversely, the analytic cohort more closely reflects standard clozapine use for schizophrenia under routine haematological monitoring. Missing laboratory data were therefore unlikely to be completely at random and may be related to treatment indication or provider-level monitoring practices. This could limit generalisability, particularly to off-label clozapine use, and may affect absolute incidence estimates, although the direction of bias cannot be determined with certainty. Moreover, this study was conducted within a single national health-care system, and the results may not be directly generalisable to settings with different monitoring thresholds, laboratory practices, or population characteristics. Second, outcome misclassification may have occurred because neutropenia was defined using WBC thresholds when ANC was unavailable. WBC-based thresholds showed high overall agreement with ANC-defined neutropenia and excellent specificity, but relatively low sensitivity, indicating that some ANC-defined neutropenia events may not have been captured. Accordingly, the use of WBC as a proxy may have led to modest underestimation of absolute incidence rates. If this misclassification was largely non-differential with respect to measured risk factors, it would be expected to attenuate associations; however, because ANC testing may vary by clinical setting or patient status, differential misclassification cannot be excluded. Importantly, sensitivity analyses restricted to ANC-defined outcomes yielded consistent results, supporting the robustness of the main findings. Third, medication exposure was defined based on prescription records rather than confirmed dispensing or adherence. However, in Taiwan's healthcare system, where medications are typically dispensed directly within hospitals or nearby pharmacies with minimal financial barriers, prescription records are likely to closely approximate actual drug exposure. Fourth, as an observational study, residual confounding by unmeasured factors—such as acute infection, nutritional status, or transient laboratory variation—cannot be excluded. Finally, we did not examine downstream clinical outcomes such as infection, relapse, or mortality, which are important considerations in balancing the risks and benefits of clozapine therapy.

Future research should build on these findings in several directions. Prospective or quasi-experimental studies are needed to evaluate whether monitoring intensity can be safely tailored based on time since initiation and individual risk profiles. Integration of clinical outcomes beyond laboratory abnormalities will be essential to determine the real-world consequences of neutropenia detection and treatment interruption.

In conclusion, this large nationwide study demonstrates that neutropenia among clozapine-treated patients is uncommon, strongly stratified by baseline risk factors, and concentrated in the early weeks of treatment. The identification of a consistent early inflection point in neutropenia risk provides empirical support for intensive early monitoring and highlights the diminishing incidence thereafter. These findings offer a robust descriptive foundation for future efforts to optimise clozapine monitoring while preserving patient safety and access to this uniquely effective treatment.

Contributors

CSW contributed to the study conception and design, methodology development, data analysis, validation, ethics approval, and drafting of the manuscript. KN contributed to the study conception and design, methodology development, and data analysis. CSW and MSW were responsible for data curation and data analysis. CCL, SHW, and WLH contributed to study conception and analytic oversight. DS contributed to study conception, methodology development, and analytic oversight. CSW and DS had final responsibility for the decision to submit for publication. All authors critically reviewed and edited the manuscript and approved the final version for submission.

Data sharing statement

The data used in this study are maintained by the Health and Welfare Data Science Center, Ministry of Health and Welfare, Taiwan. Individual-level data cannot be shared by the authors because of legal and ethical restrictions. Qualified researchers may apply to the Health and Welfare Data Science Center for access, subject to institutional review, project approval, and data-use agreements. The authors are not permitted to distribute the data directly.

Declaration of generative AI in scientific writing

During the preparation of this work, the author(s) used ChatGPT 5.2 & Grammarly in order to improve grammar and readability. After using these tools, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication's content.

Declaration of interests

DS serves on the Viatris Australian Clozapine Quality Advisory Committee and has received consulting fees for Viatris and honoraria for independent educational talks from Servier, Otsuka, Viatris, and Lundbeck. DS also reports payment for expert testimony for Davies Collison Cave Law, Member of DSMB for CLEAR (CLozapine in EARly Psychosis), and Member of Board Psychosis Australia. WLH reports research funding, without personal honoraria, from the National Health Research Institutes, the National Science and Technology Council in Taiwan, and National Taiwan University Hospital Yunlin Branch; consulting fees from Janssen, Servier, and Boehringer Ingelheim; personal honoraria for lectures from Janssen, Servier, Pfizer/Viatris, Sumitomo, Otsuka, and Boehringer Ingelheim; and travel honoraria from Orient PHARMA. CCL reports speaker's honoraria from Janssen-Cilag Pharma, Sumitomo Pharma, and Otsuka Pharma. All other authors declare no competing interests.