On average, people with schizophrenia spectrum disorders die 15 to 20 years earlier than the general population. Two-thirds of those premature deaths are from natural causes, with cardiovascular disease being the leading cause (Correll et al., 2022). This is nothing new; the mortality gap in severe mental illness (SMI) has been documented for decades, but progress in reducing it has been frustratingly slow.
Part of the problem is that the tools clinicians use to estimate cardiovascular risk were built for the general population. The Framingham Risk Score, SCORE2, and QRISK3 all rely on established risk factors such as blood pressure, lipid levels, BMI, and smoking. While these factors are highly relevant in schizophrenia, antipsychotic medications carry their own cardiometabolic consequences, and psychiatric comorbidities, psychotropic polypharmacy, and socioeconomic disadvantage may influence cardiovascular risk in ways that standard calculators do not account for (Osborn et al., 2015).
Recent work has sought to address this limitation. The Psychosis Metabolic Risk Calculator (PsyMetRiC) predicts metabolic syndrome risk in young people with psychosis (Perry et al., 2021), and has recently been expanded to predict type 2 diabetes and clinically significant weight gain (Perry et al. 2026). However, it remains focused on people aged 16-35 years and predicts cardiometabolic outcomes rather than cardiovascular disease (CVD) events. Whether cardiovascular risk can be more accurately estimated across the wider population of people with schizophrenia remains an important question.
To address this, Nielsen et al. (2026) developed a CVD risk prediction model specifically for people with schizophrenia and tested whether machine learning could improve prediction accuracy.
People with schizophrenia face a significantly higher risk of cardiovascular disease, but most cardiovascular risk prediction tools were developed for the general population and may not fully capture the factors that influence risk in this group.
Methods
The study drew on linked population-based health registers from Sweden (48,800 individuals) and Denmark (31,200), covering everyone aged 30 or over with a schizophrenia spectrum diagnosis (ICD-10: F20-F29) and no prior CVD diagnosis. Participants were followed for up to five years (2007-2019) for incident CVD events.
Three modelling approaches were compared:
- Standard logistic regression using only established CVD risk factors (hypertension, diabetes, obesity, smoking, family history).
- Lasso-penalised logistic regression using 76 predictors, including psychiatric comorbidities, psychotropic medication history, and sociodemographic variables
- XGBoost, a gradient-boosted tree algorithm that can capture non-linear interactions between predictors.
Models were developed independently in each country and then externally validated in the other country. Discrimination was assessed using the AUC (Area Under the Curve), and calibration was assessed using Brier scores and calibration plots. The study followed TRIPOD+AI reporting guidelines.
Results
Adding psychiatric and sociodemographic predictors beyond established CVD risk factors improved model performance, while more complex machine learning did not.
Model performance
- The lasso-penalised logistic regression (76 predictors) achieved the best performance in both countries: AUC of 0.745 (95% CI 0.742 to 0.749) in Sweden and 0.722 (95% CI 0.719 to 0.726) in Denmark.
- Standard logistic regression using only established risk factors achieved AUCs of 0.730 (Sweden) and 0.684 (Denmark). This is a statistically significant drop, with confidence intervals that do not overlap.
- XGBoost was in the middle at 0.734 (Sweden) and 0.704 (Denmark). While this is better than basic logistic regression, it is still not better than lasso, suggesting that additional predictors add value, but complex non-linear interactions do not.
External validation
Both models transferred well across countries. The Danish model applied to Swedish data yielded an AUC of 0.746 (95% CI 0.741 to 0.751), similar to the internal Swedish result. The Swedish model on Danish data gave an AUC of 0.720 (95% CI 0.712 to 0.726). This cross-country transportability is a meaningful finding for potential use in Europe.
Clinical thresholds
At a 10% predicted probability threshold, the Swedish model identified 67.8% of individuals who developed CVD within five years (sensitivity), with a positive predictive value of 19.0%. This indicates that roughly 1-in-5 people flagged as high risk did experience a cardiovascular event. The negative predictive value was 95.5%, suggesting the model is particularly useful for ruling out high risk.
Key predictors
Older age was the strongest single predictor. Hypertension, diabetes, obesity, and family history of CVD were the top established risk factors. Among the additional predictors, alcohol use disorder, substance use disorder, mood stabilisers, anti-epileptics, antipsychotics, and antidepressants all featured in both national models. Sociodemographic variables like income, civil status, and having children also ranked among the most important predictors.
Using linked health registry data from nearly 80,000 people with schizophrenia, researchers found that psychiatric and sociodemographic factors improved cardiovascular risk prediction more than complex machine learning methods.
Conclusions
This is the first CVD risk prediction model developed and externally validated specifically for all people with schizophrenia. The authors conclude that enriching established CVD risk factors with psychiatric comorbidities, psychotropic medication use, and sociodemographic variables improves five-year CVD prediction in this group.
Complex machine learning (XGBoost) offered no advantage over penalised logistic regression. Both models transferred between Sweden and Denmark without loss of performance. The authors argue that there is a need for validation outside Nordic countries, clinical impact studies, and model updates using directly measured cardiometabolic data.
A schizophrenia-specific cardiovascular risk model showed promising transportability across Sweden and Denmark, though further validation is needed before it can be used in routine practice.
Strengths and limitations
The scale of this study is a genuine strength. Drawing on nearly 80,000 individuals across two independent national datasets provides substantial statistical power, and the cross-country external validation addresses one of the most persistent weaknesses in clinical prediction modelling: the absence of independent replication. Many existing CVD risk models for psychiatric populations lack external validation (Osborn et al., 2015), making this a meaningful and novel step forward.
The decision to systematically compare simple logistic regression, penalised regression, and XGBoost within a single analytic framework is also a major strength. The finding that XGBoost offered no improvement over lasso regression is consistent with other literature on chronic disease prediction (Nusinovici et al., 2020) and is itself a useful contribution, challenging the pre-existing assumption that algorithmic complexity automatically improves prediction.
The most significant limitation is the reliance on registry-based proxy measures rather than directly measured clinical values. Blood pressure, BMI, and smoking are inferred from diagnosis codes and medication prescriptions, capturing only the most documented clinical presentations. This is a known issue with pharmacoepidemiological data from electronic health records. People with schizophrenia are systematically under-investigated for physical health conditions compared with the general population (Ayerbe et al., 2018), so the individuals at highest risk may also be those whose risk factors are least visible in the registers. The model may therefore underestimate risk.
The sociodemographic predictors also raise questions. Low income, being unmarried, and not having children may partly reflect structural disadvantage and inequalities in healthcare access rather than individual biological risk. The authors acknowledge this, but it warrants careful thought before clinical deployment, particularly regarding whether a tool that uses social circumstances as predictors risks compounding existing inequalities rather than addressing them.
Antipsychotics and other psychotropic medications appearing as CVD risk predictors also raise the question about interpretation. These associations may reflect the effects of the medications themselves, the severity of illness that led to their prescription, or both. The lasso identifies signals in the data, without distinguishing the drug’s direct effect from the severity of illness driving prescription. While this does not invalidate the model for prediction purposes, it limits causal interpretation.
Finally, it is worth noting that Sweden and Denmark have extraordinarily complete health records, with data from different parts of the healthcare system joined up in ways that are not the norm in other countries. Whether the model would be as accurate at prediction in the UK, where psychiatric and primary care records are less regularly linked, or in countries with fewer data resources, remains a question.
Large, externally validated datasets strengthened the model’s credibility, but the use of registry-based data and sociodemographic predictors raises important questions about accuracy, interpretation, and health inequalities.
Implications for practice
For clinicians working with people who have schizophrenia, this study reinforces current standard practice, which is that established CVD calculators likely underestimate risk in this group, and a more thorough assessment is warranted. Substance use, psychotropic medication burden, and social circumstances all exacerbate the risk, alongside blood pressure and cholesterol.
For people working with a patient with schizophrenia in a psychiatric outpatient clinic, this paper gives a clearer framework for thinking about what “cardiovascular risk” actually means for them. There needs to be an interdisciplinary approach that considers their alcohol use, anti-epileptic prescription, income, and living situation. These factors may already be in the clinical record but are often overlooked. This study quantifies the contribution of these factors to CVD risk, making the case for explicitly including them in physical health reviews.
For researchers, the most pressing next step is external validation. UK data linked to secondary care could be a candidate for this; however, the differences in how psychiatric and physical health records are linked to secondary care would need careful attention. Beyond replication, the crucial unanswered question is whether using this model changes clinical decisions and improves patient outcomes. A high AUC does not equate to clinical utility or causal inference, and that gap is wider than is often acknowledged in the prediction modelling literature. This study is a careful and rigorous step in the right direction.
Better cardiovascular risk prediction tools are only valuable if the healthcare systems and clinical infrastructure exist to act on what they tell us.
Statement of interests
Aanya Malaviya is conducting independent research on cardiovascular and metabolic outcomes in psychosis using NHS Glasgow SafeHaven electronic health records, supervised by Professor Gavin Reynolds (Sheffield Hallam University). This work overlaps in subject matter with the paper reviewed here, though she has no relationship with the authors and no other conflicts of interest to declare. AI tools were used to support the editing and reviewing of this blog.
Editor
Edited by Éimear Foley. ChatGPT assisted with language refinement and formatting during the editorial phase.
Links
Primary paper
Sara Dorthea Nielsen, Maja Dobrosavljevic, Pontus Andell, Zheng Chang, Line Katrine Harder Clemmensen, Henrik Larsson, and Michael Eriksen Benros (2026). Development and external validation of machine learning approaches for risk prediction of cardiovascular disease in individuals with schizophrenia: a nationwide Swedish and Danish study. BMJ mental health, 29(1).
Other references
Ayerbe, L., Forgnone, I., Foguet-Boreu, Q., González, E., Addo, J., & Ayis, S. (2018). Disparities in the management of cardiovascular risk factors in patients with psychiatric disorders: a systematic review and meta-analysis. Psychological medicine, 48(16), 2693-2701.
Correll, C. U., Solmi, M., Croatto, G., Schneider, L. K., Rohani‐Montez, S. C., Fairley, L., … & Tiihonen, J. (2022). Mortality in people with schizophrenia: a systematic review and meta‐analysis of relative risk and aggravating or attenuating factors. World psychiatry, 21(2), 248-271.
Nusinovici, S., Tham, Y. C., Yan, M. Y. C., Ting, D. S. W., Li, J., Sabanayagam, C., … & Cheng, C. Y. (2020). Logistic regression was as good as machine learning for predicting major chronic diseases. Journal of clinical epidemiology, 122, 56-69.
Osborn, D. P., Hardoon, S., Omar, R. Z., Holt, R. I., King, M., Larsen, J., … & Petersen, I. (2015). Cardiovascular risk prediction models for people with severe mental illness: results from the prediction and management of cardiovascular risk in people with severe mental illnesses (PRIMROSE) research program. JAMA psychiatry, 72(2), 143-151.
Perry, B. I., Osimo, E. F., Upthegrove, R., Mallikarjun, P. K., Yorke, J., Stochl, J., Perez, J., Zammit, S., Howes, O., Jones, P. B., & Khandaker, G. M. (2021). Development and external validation of the Psychosis Metabolic Risk Calculator (PsyMetRiC): a cardiometabolic risk prediction algorithm for young people with psychosis. The Lancet Psychiatry, 8(7), 589–598. https://doi.org/10.1016/S2215-0366(21)00114-0
Benjamin Perry, Emanuele Osimo, Shuqing Si, Karla Hitchins, Clara Lewis, Ben Laws, Simon Griffin, Golam Khandaker, Graham Murray, David Shiers, Carolyn Chew-Graham, Peter Jones, Alastair Denniston, Marco Bardus, Sue Jowett, Annabel Walsh, Shizana Arshad, Tomas Formanek, Toby Pillinger, Robert McCutcheon, Richard Holt, Silke Heyse, Magaly Rambousek, Khadija Whiteley, Rachel Upthegrove, Joie Ensor (2026) Cardiometabolic prediction models for young people with psychosis spectrum disorders in the UK (PsyMetRiC 2.0): a retrospective, multicohort clinical prediction model study. The Lancet Psychiatry, 13(4), 291-303.
Yanakan Logeswaran (2026) Psychosis and metabolic risk: PsyMetRiC 2.0 reaches the clinic. The Mental Elf, 26 June 2026