IPM Take
Cardiology is good at predicting how much risk a patient carries after a heart attack.
This study asks a more interesting question: risk of what?
Researchers found three markedly different five-year trajectories after myocardial infarction. Some patients accumulated predominantly cardiometabolic disease. Others developed arrhythmias, structural heart disease and kidney problems. A third group experienced progressive respiratory and multisystem deterioration and had 43.9% five-year mortality.
That could eventually move secondary prevention beyond a universal follow-up template.
But there is an important reality check. The established SMART score remained better at predicting mortality than trajectory membership alone, and adding the AI-derived trajectories improved discrimination only modestly. After full adjustment, the independent mortality association of the trajectories disappeared.
So this is not evidence that AI has beaten conventional cardiology.
Its potential value is different: telling clinicians why two patients with similar cardiovascular risk may need different follow-up care.
Executive Summary
Researchers from the University of Surrey used longitudinal health records from 12,701 UK Biobank participants with incident acute myocardial infarction to map diagnoses occurring over the following five years. Temporal machine learning identified three multimorbidity profiles.
The largest, ACUTE-CARD, included 63.4% of patients and was dominated by cardiometabolic disease with episodic cardiorespiratory complications. CARDIOMIX accounted for 13.5% and featured structural heart disease, atrial fibrillation, valvular disease and cardiorenal metabolic dysfunction. SMO-CARD, representing 23.1%, showed progressive respiratory, musculoskeletal and frailty-related disease and had the highest five-year mortality at 43.9%.
An XGBoost model predicted trajectory membership using information available before the heart attack with an AUC-ROC of 0.906. However, when diagnoses recorded during the final week before the infarction were excluded, performance fell sharply to 0.635.
The findings are therefore promising for precision follow-up, but not ready for clinical deployment.
Why it matters
- HTA bodies: Predictive accuracy alone will not establish value. Future assessment would need evidence that assigning patients to trajectories actually changes surveillance or treatment and improves outcomes beyond existing cardiovascular risk tools.
- Payers: Post-MI follow-up consumes substantial resources. Better identification of patients likely to require respiratory, renal, rhythm or cardiometabolic care could support more targeted resource allocation, but only if prospective studies show that acting on the prediction improves care.
- Industry / innovation partners: Routinely collected electronic health records could become the infrastructure for personalised secondary prevention. The opportunity is not another isolated AI score, but clinically integrated decision support that connects predicted disease trajectories to specific interventions.
Surviving a heart attack does not put every patient onto the same road.
That is the central finding from a new study published in the Journal of the American Medical Informatics Association.
Researchers analysed five years of diagnostic records from 12,701 UK Biobank participants following an acute myocardial infarction. Instead of simply counting subsequent diseases, they used temporal machine learning to examine which diagnoses appeared, in what sequence and how quickly they accumulated.
Three distinct patterns emerged.
Three patients, three very different futures
The largest profile, called ACUTE-CARD, contained 63.4% of participants. It was characterised by early clustering of hypertension, dyslipidaemia, type 2 diabetes and ischaemic heart disease, with intermittent arrhythmias and cardiorespiratory deterioration. This group had the lowest five-year mortality, at 12.8%.
A second profile, CARDIOMIX, represented 13.5%. These patients showed progressive structural heart disease, atrial fibrillation, valvular disease and cardiorenal metabolic dysfunction. Five-year mortality was 20.6%.
The most concerning trajectory was SMO-CARD, accounting for 23.1% of patients. It was characterised by COPD, respiratory failure, musculoskeletal disease and frailty-related conditions. Patients were also older and more socioeconomically deprived.
Five-year mortality reached 43.9%, more than three times the rate in the largest trajectory.
The model could also predict which trajectory a patient was likely to enter using diagnoses and demographic characteristics available before the infarction. Respiratory disease, older age and deprivation pushed predictions toward the high-risk SMO-CARD profile, while atrial fibrillation, valve disease and diabetes were important for CARDIOMIX.
That creates an obvious clinical possibility.
A patient predicted to enter CARDIOMIX might need closer rhythm, kidney and structural-heart surveillance. Someone heading toward SMO-CARD might benefit from greater attention to pulmonary disease, frailty and wider multisystem care. The authors propose precisely this kind of trajectory-specific approach.
AI added a map. It did not replace the risk score
There is, however, an important limit to the personalised-medicine story.
The established SMART cardiovascular risk score remained the stronger standalone predictor of five-year mortality.
Trajectory membership alone produced a C-index of 0.59, compared with 0.65 for SMART. Combining both increased discrimination only modestly, to 0.67. After adjustment for age, sex, deprivation, blood pressure, diabetes and previous vascular disease, the mortality effects associated with trajectory membership were no longer statistically significant.
That does not make the machine-learning approach pointless.
It changes its purpose.
SMART can tell clinicians that two patients are high risk. The trajectory model may eventually help explain why one is likely to progress through cardiometabolic disease while the other develops respiratory deterioration, frailty or cardiorenal complications.
There is another caveat.
The best-performing XGBoost model achieved an AUC-ROC of 0.906 when diagnoses recorded up to the day before the heart attack were included. When researchers removed diagnoses from the final seven days before the event to test for potential information leakage, the AUC fell to 0.635.
That suggests some of the apparent predictive power came from information generated very close to the infarction, potentially including diagnoses connected to a complicated presentation.
It is a major reason not to treat the model as clinic-ready.
The UK Biobank population also has known healthy-participant bias, was predominantly White British, and the study relied on hospital ICD-10 diagnoses. The authors call for validation in more representative populations and prospective implementation studies before clinical use.
Still, the concept is important.
Precision cardiology after myocardial infarction may not simply mean calculating increasingly accurate probabilities of another cardiovascular event.
It may mean recognising that the “heart attack survivor” is not one patient type at all.
The next step is not just predicting who is at risk. It is predicting what they are at risk of becoming, and changing care before they get there.
