Machine Learning Models for Heart Failure Hospitalisation Risk Still Perform Best on Familiar Data
Machine learning is already embedded in cardiovascular research, but its performance in heart failure risk prediction remains uneven when the target is hospitalisation rather than a cleaner laboratory endpoint. A scoping review in BMJ Open examined how investigators have used supervised models to estimate the risk of first admission, readmission, or rehospitalisation in adults living with heart failure. The review identified 27 studies and found a consistent pattern: ensemble-style approaches often outperformed simpler methods, yet the evidence base remains dominated by electronic health record variables and leaves major clinical and economic questions unresolved.
The practical problem is familiar to anyone caring for patients with heart failure. Hospitalisation is not a single biological event; it reflects congestion, renal dysfunction, arrhythmia burden, medication adherence, access to care, social support, and the speed with which early decompensation is recognized. Machine learning can sort through large clinical datasets faster than traditional regression, but it cannot correct for missing physiology or missing context. When the data are narrow, the model may be precise on paper and brittle at the bedside.
Predictive Performance Favored Ensemble Methods, but the Signal Was Modest
Across the included studies, all models used supervised learning. Most work focused on readmission rather than first-time hospitalisation, which matters because those are related but not interchangeable clinical problems. A patient returning to the hospital after a prior admission carries a different mix of risk factors, treatment exposure, and disease trajectory than a patient facing an initial admission for heart failure decompensation.
For 30-day hospitalisation or readmission risk, Extreme Gradient Boosting posted the highest mean area under the curve, at 0.69. Naïve Bayes followed at 0.68, and Deep Unified Networks at 0.66. Those values suggest only moderate discrimination. In real-world terms, they are useful for ranking risk but not strong enough to function as a stand-alone trigger for major care decisions without careful calibration and local validation.
The 90-day horizon showed stronger performance in some models. Least Absolute Shrinkage and Selection Operator and Gradient Boosting each reached a mean AUC of 0.75, with Random Forest lower at 0.67. When studies did not specify the prediction window, Categorical Boosting achieved the top mean AUC at 0.88, followed by Generalised Linear Model Net and XGBoost at 0.79. Those better numbers are encouraging, but they sit inside a heterogeneous literature with differing endpoints, variable feature sets, and unclear transportability between health systems.
One sentence captures the central tension: the best-performing models were often the ones most able to exploit patterns already present in routine records.
Electronic Health Records Dominated, Leaving Important Risk Domains Underused
Electronic health records supplied the main data source across the review. That choice makes sense operationally because EHR data are abundant, time-stamped, and relatively easy to feed into model development. Yet the same convenience narrows the clinical lens. Heart failure risk is shaped by factors that are often missing, poorly structured, or inconsistently documented in the record: symptom burden at home, medication affordability, dietary sodium exposure, transportation barriers, loneliness, health literacy, and patient-reported functional decline.
Only a minority of models incorporated patient-reported outcomes or socioeconomic variables. In a disease where readmission often follows a cascade of small failures rather than a single catastrophic one, those omissions are not trivial. Social instability can drive delayed diuretic use, missed follow-up, and late presentation with congestion. Patient-reported dyspnea, fatigue, or weight gain may appear before overt decompensation is visible in coded data. A model trained mostly on labs, diagnoses, and prior admissions may miss the first signs of trouble.
The review’s findings fit a broader pattern in cardiovascular informatics. Models built from structured health record data often perform well in retrospective testing, then lose strength when moved to a new hospital, a different payer environment, or a population with different access to care. In heart failure, where treatment pathways vary widely and comorbidity is the rule rather than the exception, generalisation is a high bar.
Hospitalisation Risk Prediction Needs More Than Accuracy Metrics
The absence of any economic evaluation is the sharpest gap in the literature summarized here. None of the studies assessed whether these tools are cost-effective. That omission matters because hospitalisation prediction is not valuable simply because it is technically sophisticated. A risk score only matters if it changes care in a way that improves outcomes, avoids preventable admissions, or uses resources more efficiently than existing clinical judgment and transitional care pathways.
In practice, an algorithm that labels thousands of patients as high risk can also overwhelm care teams if it is poorly targeted. If the alert threshold is too low, the result is alarm fatigue. If it is too high, high-risk patients are missed. Without health economic analysis, it is impossible to know whether deploying these tools would reduce avoidable admissions or simply shift workload into another part of the system.
The review also points to a methodological challenge in the field itself. AUC is useful, but it does not tell clinicians how many admissions were prevented, how many false positives were generated, or how well the model behaved across subgroups. Those are the questions that matter when a prediction tool is built to guide follow-up visits, home monitoring, telehealth escalation, or medication review.
Generalisation and External Validation Remain the Main Barriers
The review argues for stronger generalisability, broader data integration, and economic testing. That is the right sequence. Before a model can be used to direct care at scale, it must perform reliably across hospitals, regions, and patient subgroups. It also needs input variables that reflect the true biology and lived experience of heart failure rather than the subset of information most easily captured in a database.
There is also a mechanistic reason ensemble methods may be performing better in this space. Heart failure admissions are driven by nonlinear interactions between renal function, volume status, age, prior events, medication patterns, and comorbidity. Methods such as XGBoost, Gradient Boosting, and Random Forest are designed to capture those interactions more flexibly than simpler approaches. Even so, statistical flexibility cannot compensate for incomplete inputs. If important drivers are absent, the model may simply become better at learning the wrong signals.
The field now faces a more demanding standard than improved discrimination alone. Future work will need external validation, better representation of patient-reported and socioeconomic factors, and a realistic assessment of whether prediction tools reduce hospital use without adding cost or complexity. For now, the evidence suggests potential, but not readiness for broad clinical adoption without further testing.