Predicting cardiovascular death in overweight/obese people with prediabetes using machine learning—a proof-of-concept study

Abstract

Background

Individuals with prediabetes face an increased risk of cardiovascular (CV) complications, which can ultimately lead to premature mortality. However, existing risk stratification tools are not targeted for people with prediabetes. We aimed to develop a simple explainable model to predict if a person will develop a fatal CV outcome or not among people with prediabetes.

Methods

Participants ≥ 45 years with prediabetes (HbA1c 39–47 mmol/mol (5.7–6.4%)) and established CV disease and overweight/obesity were included. A binary logistic regression model was trained to predict CV death using stratified threefold cross-validation. The model’s risk estimates were calibrated, and the predictive capability was evaluated using receiver operating characteristic (ROC) area under the curve (AUC), precision and recall.

Results

In total 5636 participants with 182 (3.2%) CV deaths (mean time-to-event of 2.0 years) and a mean trial duration of 3.3 years were included. Seven easily collected demographic and clinical variables were selected for the model. Discrimination (ROC AUC) was acceptable at 0.730 (95% CI 0.659–0.801). Applying our prediabetes cohort on existing benchmark models developed for major adverse cardiovascular events (MACE) in a general and type 2 diabetes population without prior CVD, demonstrated lower performance for CV death (ROC AUC: 0.630 (SCORE2) and 0.643 (SCORE2-Diabetes)) compared to our model. Lower performance was also observed for predicting MACE in our cohort (ROC AUC: 0.596 and 0.603) using the established models compared to the original populations (ROC AUC: 0.739 and 0.66–0.73). No comparative models for people with prediabetes and prior CVD exists. Thus, even with the limitations in different populations and outcome targets, this indicates that prediabetes-specific prediction models could potentially improve early prevention in this high-risk population.

Conclusion

We have developed a prediabetes-specific proof-of-concept model that predicts whether a person is at high risk of cardiovascular death. External validation of the model is crucial before adoption to a real-world setting to clarify whether the model generalizes beyond the studied population.

Graphical abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s12933-026-03210-3.

Research insights

What is currently known about this topic?

• Individuals with prediabetes face an increased risk of cardiovascular (CV) complications, but existing risk stratification tools are not targeted for this population.

What is the key research question?

• Can risk prediction of fatal CV outcomes for individuals with prediabetes be improved by developing prediabetes-specific prediction models?

What is new?

• Using seven easily collected demographic and clinical variables, the classification model predicts a person’s short-term risk of CV death with acceptable performance. The model demonstrated improved performance when compared to two benchmark models (developed for a general and Type 2 Diabetes population, respectively) evaluated against the same prediabetes population.

How might this study influence clinical practice?

• The model holds the potential to improve identification of individuals with prediabetes at highest risk of fatal CV outcomes.

Background

Micro- and macrovascular complications and mortality impose a major burden for individuals with prediabetes [1–4]. People with HbA1c-defined prediabetes have an 10–20% increased risk of cardiovascular disease (CVD) and mortality compared to the general population [5]. Additionally, it is found that people with prediabetes have a higher 12-months risk of major adverse cardiovascular events (MACE) compared with people having type 2 diabetes, indicating a need for management of cardiovascular (CV) complications before people with prediabetes potentially progress to type 2 diabetes [6]. Prediabetes is a metabolic condition that is characterized by blood glucose levels that are higher than normal, but below the threshold for type 2 diabetes. The condition affects millions of people worldwide and the prevalence is projected to increase looking forward [7, 8]. The diagnostic criterion for prediabetes based on HbA1c is 5.7–6.4% (39–47 mmol/mol) according to the American Diabetes Association (ADA) and 6.0–6.4% (42–47 mmol/mol) according to the International Expert Committee (IEC). In addition, impaired fasting glucose (IFG) and impaired glucose tolerance (IGT) are also used as diagnostic criteria for prediabetes [9, 10]. The prediabetes prevalence was estimated to 38% in the adult U.S. population in 2017–2020 based on IFG and HbA1c [11].

As people with prediabetes are estimated to constitute a rather large group at increased risk of CVD and mortality, it is important to focus on how to prevent or delay the onset of these complications [4, 12, 13]. A key element is identification of people at highest risk of CV events, ultimately resulting in CV mortality. Early identification of people at high risk allows for targeted prevention strategies, such as lifestyle changes or pharmacological interventions. Furthermore, understanding key risk factors is essential in this task, which highlights the importance of exploring how specific clinical variables impacts the risk. This is in line with EASD and ADA suggesting that people with prediabetes should be informed about their risk of complications and introduced to strategies to lower the risk [14, 15].

Currently, risk stratification tools exist to the general population and diabetes population [16–18]. However, these tools do not perform well when applied to people with prediabetes, indicating the need for establishment of prediabetes-specific risk score tools [13, 19]. Only one study has investigated development of a risk prediction model specifically for individuals with prediabetes [20]. It showed moderate and acceptable results for machine learning models to predict CV complications among people with prediabetes, but it did not focus specifically on CV death. Moreover, the included population had no history of previous CV complications, and the model included more than 100 clinical predictors, compromising clinical utility.

In this study, CV death events were chosen as the endpoint, since death is the ultimate consequence of CV related complications and since all included subjects already had history of CVD. The primary aim was to develop a simple explainable binary classification model to predict whether a fatal CV event will occur in short term among people with prediabetes having preexisting CVD. Secondarily, to explore which clinical information were most associated with an elevated risk of CV death.

Methods

Source of data

This study was based on individual data from a randomized controlled clinical trial, which investigated the effects of semaglutide on cardiovascular outcomes in patients with overweight and obesity without history of diabetes (SELECT) [21, 22]. The trial consisted of 17,604 subjects with preexisting CV disease, an age ≥ 45 years and with a BMI ≥ 27 kg/m2. The preexisting CV disease was defined as at least one of the following: prior myocardial infarction, prior stroke or symptomatic peripheral arterial disease. Furthermore, the study population were randomized in a 1:1 ratio with the intervention group receiving semaglutide subcutaneously 2.4 mg once-weekly and the control group receiving placebo. Primary endpoints were nonfatal myocardial infarction, nonfatal stroke and death from cardiovascular causes (3-point MACE) measured by time from randomisation to first occurrence of event. A subgroup of participants from the SELECT trial were included in this study based on the eligibility criteria: 1) being in the placebo group, 2) having prediabetes at baseline defined by an HbA1c level between 39–47 mmol/mol (5.7–6.4%) and 3) having completed the trial. Thus, subjects receiving semaglutide and subjects being in the normoglycemia range at baseline (HbA1c < 39 mmol/mol (< 5.7%)) were excluded.

Features & outcome

Clinical information collected at baseline were extracted for each participant (Additional Table 1). This included demographical variables, vital signs, clinical laboratory tests, patient reported outcomes, concomitant medication and medical history of cardiovascular events and comorbidities. Clinical laboratory tests encompassed information about biochemistry, haematology, lipids and urine analysis. Concomitant medication included cardiovascular related medication, chronic kidney disease related medication, weight management related medication and glucose lowering medication. Beside this, information on whether or not the participant had a CV death event reported within the trial period were extracted as a binary variable for each subject, defined as the outcome label.

Data pre-processing

Prior to data processing, the full dataset was split into a training and test dataset in a 70:30 ratio stratified by the outcome label, ensuring a representative distribution of participants with an event in each subset. The data processing included handling missing values and outliers, encoding of categorical features and normalization, to ensure that the extracted variables were formatted correctly for the machine learning framework.

All extracted features were categorised into three groups (numerical features, ordinal categorical features, and nominal features) to be processed differently. Initially, features with more than 10% missing values were removed. In addition, features having zero variance were excluded, as they do not contribute with any information. To handle the missing values three different imputation methods were applied. Mean imputation was used to impute missing values for numerical features, mode imputation was used for ordinal categorical features, and a new "unknown" category was created for nominal categorical features. For encoding of the categorical features, ordinal categorical variables were transformed using label encoding to maintain their inherent order, whereas one-hot encoding was applied to the nominal categorical features. Lastly, normalization was applied to scale the features within the dataset. Z-score normalization was applied on the numerical features since this is less sensitive to outliers and min–max (0–1) scaling was applied on the categorical features. The pre-processing was initially conducted solely on the training data ensuring that no information from the test set influenced the training phase, thus preventing data leakage. Afterwards, the test data was processed using the parameters derived from the training data.

Machine learning model & model explainability

A binary logistic regression model was trained to predict CV death. The model was trained using stratified threefold cross-validation on the training dataset with hyperparameters optimized through grid search. The parameter grid included regularization, class weight, penalty type, solver, tolerance and maximum iteration. Prior to this, forward feature selection was used to identify a feature subset for the model, that provided the best predictive performance based on the receiver operating characteristic curve (ROC) area under the curve (AUC). Feature selection was stopped when the ROC AUC performance improvement fell below 0.005. The final model’s performance was evaluated on the separate test dataset utilizing the following metrics: ROC AUC, precision, recall and accuracy. In addition, the predicted probabilities were plotted in a histogram along with the actual outcome to visually demonstrate the model’s ability and confidence in separating participants with and without a CV death event. Lastly, the coefficients from the logistic regression model were used to interpret the feature’s influence on the predicted risk and a physiological evaluation was used to validate the identified predictors and patterns from a clinical perspective.

Moreover, the model’s performance was compared against two models (SCORE2 [23] and SCORE2-Diabetes [18]), which are recommended for CVD risk prediction in clinical guidelines for the general population and people with type 2 diabetes, respectively [24, 25]. The clinical variables used in these models were sex, age, nationality, smoking status, systolic blood pressure, total cholesterol, HDL-cholesterol, HbA1c and estimated glomerular filtration rate (eGFR). A detailed description of the SCORE2 and SCORE2-Diabetes used to evaluate these models on participants with prediabetes in this study can be found in Additional methods 2. There is currently not a prediabetes-specific model publicly available, to test our cohort in an existing prediabetes model. Lastly, model calibration (Additional methods 1) was applied on top of the logistic regression model to adjust the predicted probabilities to better reflect the true observed number of events. Brier score and log loss were used to evaluate the calibration. Further, risk stratification groups were defined as: low (0–2.5%), moderate (2.5–5%), high (5–10%), and very high (> 10%).

Data processing and model development was implemented in Python (version 3.11.6). Further, the study followed the TRIPOD + AI (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis + Artificial Intelligence) guidelines for reporting prediction model studies [26].

Results

A total of 5636 participants were included in this study. The baseline characteristics for the participants are shown in Table 1. The participants represented a global population representing 41 countries (Additional Figs. 1–2). In total 3.2% (n = 182) of the participants had a CV death event, 8.2% (n = 464) had a 3-point MACE event and 2.4% had a non-CV death event. The participants had a mean age of 62.1 years with a mean BMI at 33.6. Furthermore, 72.5% of the population were men. Additionally, the medical history of cardiovascular events and comorbidities for the participants are shown in Table 2. The trial duration ranged between 26 days and 4.6 years (1680 days) with a mean trial duration of 3.3 years (1222 days). The mean time to CV death event was 732 days (2.0 years) ranging between 26 days and 4.33 years (1582 days). In total, 2.8% of the participants with baseline HbA1c level between 5.7–5.9% experienced a CV death event and 3.6% of the participants with baseline HbA1c level between 6.0–6.4% experienced a CV death event. Furthermore, no apparent differences between the participants in the training and test data was seen (Table 1).

Table 1.

Baseline characteristics of participants. Mean ± standard deviation is shown for the total population and for the train and test data, respectively.

	Total (n = 5636 (100%))	Train (n = 3945 (70%))	Test (n = 1691 (30%))
Characteristic	Mean ± std	Mean ± std	Mean ± std
Sex
Males	4087 (72.5%)	2876 (72.9%)	1211 (71.6%)
Females	1549 (27.5%)	1069 (27.1%)	480 (28.4%)
Age, years	62.1 ± 8.6	62.1 ± 8.69	62.1 ± 8.55
HbA1c, %	5.97 ± 0.21	5.97 ± 0.21	5.96 ± 0.20
HbA1c: 5.7–5.9% (39–41 mmol/mol)	2918 (51.8%)	2023 (51.3%)	895 (52.9%)
HbA1c: 6.0–6.4% (42–47 mmol/mol)	2718 (48.2%)	1922 (48.7%)	796 (47.1%)
BMI, kg/m2	33.6 ± 5.1	33.61 ± 5.16	33.51 ± 4.87
Race (no (%))
White	4708 (83.5%)	3281 (83.2%)	1427 (84.4%)
Asian	500 (8.9%)	349 (8.8%)	151 (8.9%)
Black or African American	190 (3.4%)	150 (3.8%)	40 (2.4%)
Other	185 (3.3%)	129 (3.3%)	56 (3.3%)
Not reported	53 (0.9%)	36 (0.9%)	17 (1.0%)
Ethnic group (no (%))
Hispanic or Latino	516 (9.2%)	380 (9.6%)	136 (8.0%)
Not Hispanic or Latino	5066 (89.9%)	3528 (89.4%)	1538 (91.0%)
Not reported	54 (0.9%)	37 (0.9%)	17 (0.9%)
Body weigth, kg	97.20 ± 17.98	97.30 ± 18.17	96.99 ± 17.51
Waist circumference, cm	111.9 ± 13.1	111.92 ± 13.24	111.84 ± 12.78
Median high-sensitivity CRP level, mg/L (IQR)	1.92 (0.90 – 4.24)	1.89 (0.90 – 4.26)	1.97 (0.92 – 4.21)
eGFR, mL/min/1.73m2	81.95 ± 17.18	81.80 ± 17.25	82.29 ± 16.99
Median lipid level, mmol/L (IQR)
Total cholesterol	3.95 (3.40 – 4.72)	3.97 (3.41 – 4.71)	3.92 (3.39 – 4.76)
HDL cholesterol	1.12 (0.96 – 1.32)	1.12 (0.96 – 1.31)	1.12 (0.96 – 1.32)
LDL cholesterol	2.01 (1.58 – 2.63)	2.01 (1.58 – 2.61)	2.02 (1.55 – 2.69)
Triglycerides	1.56 (1.16 – 2.19)	1.57 (1.16 – 2.21)	1.53 (1.15 – 2.12)
Systolic blood pressure, mmHg	131.2 ± 15.2	131.39 ± 15.23	130.68 ± 14.98
Diastolic blood pressure, mmHg	79.1 ± 9.8	79.24 ± 9.80	78.87 ± 9.85
Pulse, beats/min	68.8 ± 10.7	68.88 ± 10.73	68.54 ± 10.51
EQ-5D-5L index score	0.88 ± 0.15	0.87 ± 0.15	0.88 ± 0.14
EQ-5D-VAS score	76.77 ± 15.75	76.53 ± 16.08	77.34 ± 14.95
CV death event *
Yes	182 (3.2%)	126 (3.2%)	56 (3.3%)
No	5483 (96.8%)	3819 (96.8%)	1635 (96.7%)
Cause of CV death
Sudden cardiac death	80 (44.0%)	57 (45.2%)	23 (41.1%)
Acute myocardial infarction	10 (5.5%)	6 (4.8%)	4 (7.1%)
Stroke	17 (9.3)	10 (7.9%)	7 (12.5%)
Heart failure	10 (5.5%)	4 (3.2%)	6 (10.7%)
Cardiovascular procedure	4 (2.2%)	3 (2.4%)	1 (1.8%)
Undetermined	55 (30.2%)	40 (31.7%)	15 (26.8%)
Other	6 (3.3%)	6 (4.8%)	0 (0.0%)
3-point MACE events **	464	314	150
Cardiovascular death	153 (32.9%)	107 (34.1%)	46 (30.7%)
Non-fatal acute myocardial infarction	210 (45.3%)	136 (43.3%)	74 (49.3%)
Non-fatal stroke	101 (21.8%)	71 (22.6%)	30 (20.0%)

*Defined by time-to-first CV death event

**Defined by time-to-first occurrence of 3-point MACE event

Table 2.

Medical history of cardiovascular events and comorbidities of the included population (n = 5636)

	Count (%)
History of cardiovascular events
Carotid artery stenosis	274 (4.9%)
Carotid revascularization	129 (2.3%)
Chronic heart failure	1431 (25.8%)
Coronary artery stenosis	3229 (57.3%)
Coronary heart disease	4678 (83.0%)
Coronary revascularisation	3901 (69.2%)
Echocardiography	3777 (67.0%)
Hypertension	4644 (82.4%)
Left ventricular ejection	653 (11.8%)
Myocardial infarction	4355 (77.3%)
Peripheral artery stenosis	268 (4.8%)
Stroke	1229 (22.2%)
Symptomatic peripheral arterial disease	527 (9.5%)
Transient ischaemic attack	231 (4.2%)
History of comorbidities
Asthma	392 (7.1%)
Chronic kidney disease	652 (11.8%)
Chronic obstructive pulmonary disease	516 (9.3%)
Gout	502 (9.1%)
Hips osteoarthritis	418 (7.6%)
Knee osteoarthritis	912 (16.5%)
Non-alcoholic fatty liver	508 (9.2%)
Non-alcoholic steatohepatitis	39 (0.7%)
Sleep apnoea syndrome	772 (13.9%)
Trial inclusion criteria (preexisting CV disease)
Only myocardial infarction (MI)	3870 (68.7%)
Only stroke	925 (16.4%)
Only Peripheral arterial disease (PAD)	268 (4.8%)
MI + stroke	207 (3.7%)
MI + PAD	179 (3.2%)
Stroke + PAD	47 (0.8%)
MI + stroke + PAD	23 (0.4%)
Other	117 (2.1%)

Data pre-processing

A total of 179 baseline features were extracted from the SELECT trial, including 12 demographical features, 12 vital sign features, 43 clinical laboratory features, 18 patient reported outcomes, 62 medication features and 32 features related to medical history of cardiovascular events and comorbidities. A full list of baseline variables can be found in Additional Table 1. Pre-processing of the 179 extracted features resulted in four features being excluded for more than 10% missing data (Urine HCG, Creatine Kinase in serum, Amylase in serum, Lipase in serum) and 22 features being excluded due to zero variance (Additional Table 2). Parameters used for imputation and normalization can be found in Additional Table 3. Further, one hot encoding of the nominal categorical features increased the number of features by 23, thus resulting in 177 features being available for the machine learning model.

Model performance

The model performance demonstrated a ROC AUC of 0.730 (95% CI 0.659–0.801) (Fig. 1). ROC curves for training and validation can be found in Additional Fig. 6. The model’s ability to classify non-cases from cases using a decision threshold at 0.5 on the uncalibrated estimates, demonstrated a recall at 0.589 (95% CI 0.459–0.708), precision at 0.068 (95% CI 0.049–0.095) and accuracy at 0.721 (95% CI 0.699–0.742) (Table 3). The mean predicted probability for participants experiencing a CV death event was higher than the risk estimates for participants not experiencing a CV death (Additional Figs. 3, 4, and 5). Confusion matrices showing the model’s predictions compared with the true observations for training, validation and test, respectively, can be found in Additional Fig. 7.

Fig. 1.

ROC curve for the prediabetes-specific model presented in this study and ROC curves for SCORE2 and SCORE2-Diabetes when utilizing participants with prediabetes from this study

Table 3.

Performance metrics for the developed model during training, validation and test, respectively

	ROC AUC [95% CI]	Precision [95% CI]	Recall [95% CI]	Accuracy [95% CI]
Training	0.775 [0.737, 0.814]	0.073 [0.059, 0.091]	0.603 [0.516, 0.684]	0.743 [0.729, 0.757]
Validation	0.762 [0.721, 0.803]	0.076 [0.062, 0.094]	0.635 [0.548, 0.714]	0.742 [0.728, 0.755]
Test	0.730 [0.659, 0.801]	0.068 [0.049, 0.095]	0.589 [0.459, 0.708]	0.721 [0.699, 0.742]

When utilizing SCORE2 and SCORE2-Diabetes on participants with prediabetes from this study, the performance demonstrated a ROC-AUC of 0.630 and 0.643 for CV death (Fig. 1) and 0.596 and 0.603 for 3-point MACE, respectively. The performance from the original SCORE2 and SCORE2-Diabetes when applied on a general population and a type 2 diabetes population, were a ROC AUC of 0.739 and 0.66–0.73, respectively.

Calibration of the predicted probabilities improved the brier score and log loss from 0.201 and 0.598 to 0.032 and 0.137, respectively (Additional methods 1). Thus, the calibrated model did more accurately reflect the true likelihood of an event, making it more clinically interpretable and relevant for a real-world application. The complete calculation of short-term risk of CV death using the presented model, can be found in Additional Table 5. For example, the estimated risk of CV death was 5.8% (risk group: high) for a 56-year-old man with history of CHF, hsCRP of 3.79 mg/L, gamma glutamyl transferase of 10 U/L, ALT of 10 U/L and level of leucocytes in blood of 12.07 10^9/L (Additional Fig. 8 and Additional Table 4). For a woman with the same clinical characteristics the estimated risk was 4.46% (risk group: moderate).

Converting the estimated risk scores for all test subjects into the four defined risk categories, showed that among participants who experienced a CV death event, 19.6% were categorised as low risk, 42.9% as moderate risk, 25.0% as high risk, and 12.5% as very high risk (Fig. 2). Among participants not experiencing an event, the majority (53.1%) was categorised as low risk.

Fig. 2.

Risk stratification of the test subjects using the presented model. The height of the bars represents the percentage of participants in each risk group within each of the two groups (participants experiencing a CV death event and participants not experiencing a CV death event, respectively). The number of participants in each bar is represented by n

Model features

A total of seven features were identified from the forward feature selection as the best feature subset for the logistic regression model. The selected features were sex, age, history of chronic heart failure event, high sensitive C-Reactive Protein in serum (hsCRP), gamma glutamyl transferase (GGT) in serum, alanine aminotransferase (ALT) in serum and leucocytes in blood (Table 4). The logistic regression model was trained using the following hyperparameters: regularization: 0.005, class weight: ‘balanced’, max iterations: 100, penalty: ‘l2’, solver: ‘saga’, tolerance: 1e-06.

Table 4.

Features selected for the logistic regression model including model coefficient and clinical interpretation

Feature & description	Feature category	Model coefficient	Clinical interpretation
Medical history of chronic heart failure event (0 = no event, 1 = event)	Cardiovascular history	+ 0.7057	Heart Having a previous chronic heart failure event was associated with an increase in the estimated CV death risk
High Sensitive C-Reactive Protein (hsCRP) in Serum (mg/L)	Laboratory results (biochemistry)	+ 0.3171	Inflammation Higher level of high sensitive C-Reactive Protein was associated with an increased estimated CV death risk
Age (years)	Demographics	+ 0.3100	Age Higher age was associated with an increase in the estimated CV death risk
Alanine Aminotransferase (ALT) in serum (U/L)	Laboratory results (biochemistry)	−0.2747	Liver Higher level of Alanine Aminotransferase was associated with a decreased estimated CV death risk
Sex (0 = female, 1 = male)	Demographics	+ 0.2508	Gender Male gender was associated with an increased estimated CV death risk
Gamma glutamyl transferase (GGT) in serum (U/L)	Laboratory results (biochemistry)	+ 0.1849	Liver Higher level of gamma glutamyl transferase was associated with an increase in the estimated CV death risk
Leucocytes in blood (10^9/L)	Laboratory results (haematology)	+ 0.1413	Blood Higher level of leucocytes in blood was associated with an increase in the estimated CV death risk

Among the seven features in the model, the only inverse correlation was found between ALT and CV death. The mean ALT level at baseline was 22.9 U/L for participants getting a CV death event and 27.7 U/L for participants not getting a CV death event (difference: -4.8 U/L). Further, the mean level of ALT was analysed for different age groups (Fig. 3). It was found that the mean level of ALT was lower for the CV death groups compared to the no CV death groups for all age groups. The differences between the two groups ranged between 8.2 U/L to 0.4U/L.

Fig. 3.

Mean level of alanine aminotransferase in serum (ALT) for different age groups divided by CV death status. The y-axis is centered around the mean level of ALT for all participants. Age groups are divided into 5-year bins with number of participants within the bin denoted by n. Delta show the difference (U/L) between the two groups within each age group. The error bars show one standard deviation. Participants with missing ALT level (n = 97) was not included in this analysis

Discussion

This study presented a binary classification model to predict the short-term risk of CV death in overweight/obese people with prediabetes having preexisting CVD. People with prediabetes is a clinically relevant population since cardiovascular risk is already elevated at this stage, but preventive strategies are often less applied than in people with type 2 diabetes. Moreover, existing cardiovascular risk models are typically developed for either the general population or people with type 2 diabetes, thus leaving a gap in risk stratification in between that are targeted for people with prediabetes. The model was well-calibrated and demonstrated an acceptable level of discrimination, showing improved performance when compared to two benchmark models evaluated against this prediabetes cohort. The model included seven easily collected demographic and clinical variables and the features’ associations with the estimated CV death risk were provided. This showed the possibility of using a simple machine learning framework to identify high risk individuals for CV death while maintaining interpretable results, which is crucial for clinical utility. However, for a real-world clinical application the model’s performance needs to be externally validated.

The performance of our model on CV death demonstrated a ROC AUC of 0.730. Testing two established CVD risk scores (SCORE-2 and SCORE2-Diabetes) on the same prediabetes population, achieved an AUC of 0.630 and 0.643. Thus, our model demonstrated improved absolute AUC performance of + 0.10 (15.9% improvement) and + 0.087 (13.5% improvement) compared to these benchmark models. However, some limitations must be taken into consideration, when comparing the model in this study with SCORE2 and SCORE2-Diabetes. The populations used in the two established models do not include people with a history of CVD and they use a 10-year horizon compared to a trial duration less than 5 years in this study [18, 23]. Furthermore, 3-point MACE is the primary endpoint in SCORE2 and SCORE2-Diabetes, thus evaluating these models on CV death as endpoint might be difficult. Therefore, 3-point MACE performance were also evaluated using the prediabetes population in this study. This showed a decreased ROC AUC performance of -0.143 (−19.3%) (SCORE2) and −0.063 (−8.6%) to −0.127 (−17.4%) (SCORE2-Diabetes), when this prediabetes cohort was compared to the original populations. These results indicate a potential benefit from developing CV risk models specifically targeted to people with prediabetes instead of using the established models targeted for the general population or people with type 2 diabetes. Because there are no publicly available CVD risk models specifically for prediabetes, we could not compare our cohort’s performance against a prediabetes‑specific model, which would have provided a more direct comparison.

The model performance from this study was comparable to existing literature focusing on predictive models in similar patient populations and outcomes. A study developed 5-year micro- and macrovascular prediction models in the prediabetes population using logistic regression and gradient-boosted decision trees [20]. The study showed comparable performance between the two machine learning models with a ROC AUC at 0.73 for peripheral vascular disease, 0.69 for cerebrovascular disease and 0.71 for CVD [20]. However, the study population in [20] was based on a large number of predictors and represented only one nationality, not including information about race and ethnicity, thus limiting the model’s clinical utility and generalizability. In addition, people with a history of previous CV complications were excluded, indicating a lower risk population compared to the population in this study. To the authors knowledge, no study has previously focused on development of a prediction model specifically for CV death among people with prediabetes and prior CVD. The motivation for this proof-of-concept study was therefore to explore whether a discriminative machine‑learning approach could capture patterns associated with cardiovascular mortality in a high‑risk subgroup of individuals with prediabetes. A few studies focused on CV death among people with type 2 diabetes, showing performance between 0.78–0.84 [27–29]. However, none of these studies validated the model on an independent test dataset, which could explain part of the better performance. Furthermore, several studies reported AUC values ranging from 0.70 to 0.80 when predicting several CV outcomes in a type 2 diabetes population. For instance, [30] showed a mean ROC AUC at 0.70 (MI: 0.70, stroke: 0.72, CVD: 0.68) using a 5-year logistic regression model and [31] showed a performance of 0.70 in predicting the 4-year risk of non-fatal or fatal MI, stroke or cardiovascular death. [32] showed a ROC AUC for MACE at 0.73 (MI: 0.72, stroke: 0.70, heart failure: 0.73) in a type 2 diabetes population receiving SGLT2 inhibitor therapy. Furthermore, a review has evaluated the performance of 87 machine learning models predicting micro- and macrovascular complications among adults with type 2 diabetes [33]. It was found that macrovascular outcomes in general had lower performance than microvascular outcomes. Further, the majority (46%) had a performance between 0.6 and 0.75 interpreted as “possibly helpful”. 18% had a poor performance below 0.6 and 36% had a performance above 0.75, interpreted as a “clearly useful discrimination”.

The selected features in the model consisted of two demographic variables, one CV history variable and four variables extracted from clinical laboratory tests, which were clinically related to heart function, inflammation, liver, blood, gender and age. Age, male gender, history of chronic heart failure, hsCRP, gamma glutamyl transferase and level of leucocytes, were all found to be positively associated with CV death. It was observed that history of chronic heart failure (CHF) contributed with the highest positive coefficient in the model, indicating that impaired cardiac function had highest impact on an increased risk of CV death. This correlation aligns with existing literature showing a lower survival rate among people with CHF [34, 35]. Further, a positive correlation was found for hsCRP, suggesting that higher levels of inflammation in the body increased the CV death risk. This also aligns with existing literature showing a positive correlation between increased level of hsCRP and CVD and mortality [36, 37]. Further, the level of gamma glutamyl transferase and leucocytes were selected among the best predictive features. Gamma glutamyl transferase level is clinically related to liver function, while the level of leukocytes in the blood is associated with the immune system. Both features showed a positive correlation to CV death, meaning that increased gamma glutamyl transferase and leucocytes, independently, increased the estimated risk of CV death. Lastly, higher age and male gender increased the risk. A positive age and gender correlation were also found in [20] in predicting CVD among people with prediabetes. Of interest, ALT showed an inverse correlation, indicating that higher circulating ALT concentrations decreased the risk of CV death. This inverse correlation between ALT and CV death might be counterintuitive as elevated levels are typically associated with liver diseases. However, studies regarding the relationship between ALT and mortality show contradictory findings. For example, a positive correlation was found in [38], whereas a negative correlation was found in [39] and [40]. Since it is expected that the level of alanine aminotransferase decreases with age, the mean level of alanine aminotransferase was analysed for different age groups in this study. This analysis showed that participants experiencing CV death had lower levels of ALT at baseline compared to those in the same age group who did not experience an event.

Surprisingly, clinically actionable features like HbA1c, BMI or cholesterol, which has often been used in similar models (e.g. in SCORE2 and SCORE2-diabetes), were not selected in this model. This could be explained by the relatively low HbA1c variability ranging from 39–47 mmol/mol (5.7–6.4%) in the study. However, a higher mortality rate was observed among people in the higher prediabetes range at baseline (3.6% CV death events) compared to the lower prediabetes range (2.8% CV death events). Further, lower variability was also seen for BMI, since all included participants were categorised as overweight or obesity (BMI ≥ 27 kg/m2), thus not representing normal or underweight BMI levels. Typically, prioritization of actionable features is more often seen in traditional regression models focussing on estimating the absolute risk of an event. In contrast, this work focussed on an exploratory and data-driven modelling approach in which the model selected the feature combination resulting in the best model performance, independent of whether the features were clinically modifiable. In this context, it is important to mention that features not selected in the model did not mean zero correlation with CV death, but adding more features did not improve the discriminative performance substantially compared to the potential risk of overfitting. However, actionable predictors (e.g. smoking, systolic blood pressure, total cholesterol, HDL-cholesterol, HbA1c) can be clinically useful because they point directly to interventions. Thus, subsequent work could explore models that prioritize modifiable features even if this requires a trade‑off in predictive performance to increase clinical utility.

Among the limitations of this study, is the low precision of the model. This resulted in a relatively high risk of false positives, emphasizing the need to explore if improving the model’s performance is possible. We selected CV death as the outcome because it represents the most definitive consequence of CV complications. Furthermore, since all participants had a history of CV disease, evaluating fatal CV events as the subsequent outcome is particularly relevant. However, a potential consideration for further model development could be to focus on a broader outcome, such as 3-point MACE (CV death, non-fatal myocardial infarction and non-fatal stroke). This would increase the number of cases and at the same time be a relevant endpoint from a clinical point of view and align the model with commonly used clinical outcomes. In addition, future work could focus on applying more sophisticated non-linear modelling and potentially including a larger variety of clinical variables, e.g., repeated measures (time-series) of clinical and laboratory variables to capture temporal changes in cardiovascular risk. However, while complex and non-linear learning models may demonstrate a higher performance, the clinical explainability of simple machine learning models, provides important clinical value. Furthermore, the proposed logistic regression model treats CV death as a binary outcome, meaning that time-to-event information was not used. Our primary clinical use case was whether a person with prediabetes at time of consultation belongs to a high‑risk group who may benefit from further attention or intensified management, rather than the exact timing of a potential event. In this context, classification models remain clinically useful despite not explicitly modelling time‑to‑event. Regarding competing events, non-CV deaths were added to the negative class (no CV death), which is a conservative approach possibly resulting in an underestimation of the ROC-AUC as these patients might have had high risk for a CV death event. Furthermore, we assessed that the bias regarding censoring and competing events was limited since only completed subjects were included, and a low number of non-CV deaths were observed. However, time‑specific absolute risks and handling of censoring and competing events could be obtained by re‑analysing the data using survival or competing‑risk models (e.g., Cox and Fine‑Gray models). Additionally, penalized survival regression, such as LASSO-Cox, could be considered for feature selection and handling of overfitting in time-to-event analyses.

A key strength of this study was the utilization of data from a large randomized clinical trial. This enabled the possibility of including an exhaustive feature dataset, incorporating demographic information, physical measures, clinical laboratory results, medication usage, and medical history, allowing for a comprehensive exploration of the most important clinical features affecting the risk of CV death. Additionally, the SELECT trial included a global population, including diversity in geography, ethnicity, gender and age. This strengthened the results and increased the ability to generalize the results to a wide population. Some limitations should also be noted regarding the included cohort. The study was based on overweight or obese individuals above 45 years with established CV disease. These characteristics at baseline characterizes the included population as a high-risk group for subsequent CV events, thus the CV death event rate in this study (3.2% from a mean trial duration of ~ 3 years) is expected to be higher than in the general prediabetes population. As comparison, another study included 4,511 subjects with prediabetes from the NHANES dataset, where 180 CV deaths (3.99%) were observed during a median follow-up of 100 months (~ 8.33 years) [41]. Approximating total person‑years as 4,511 × 8.33 = 37,500, the observed CV death rate was approximately 0.48% per year (180 / 37,500 ≈ 0.0048 per person‑year), which corresponds a 3-year cumulative CV death risk of 1.4% (1 − exp(− 3 × 0.0048) ≈ 1.4%) and a 5‑year risk of 2.4%. Therefore, the model’s generalizability to e.g. lower-risk or younger populations needs further investigation. Additionally, the majority of the study population were males (72.5%), which could have induced bias to the model. Besides, the controlled nature of a clinical trial may limit the generalizability to a real-world setting. Thus, a key element of future work will be external validation of the model using real-world data, which better reflect a heterogenous clinical setting. This will be a crucial step before the model can be implemented in a real-world setting to assist clinicians in providing better personalized guidance for early prevention of CV complications.

Another strength in this study was the model evaluation on an independent test set. This test dataset remained completely unused during both processing and model training, showing generalizability to unseen data. Lastly, the simple machine learning framework, allowed for high model explainability, which enhanced the applicability of the model in a clinical context. Only seven features, which can easily be obtained in clinical practice, are needed as input to the model. Furthermore, the risk stratification, categorizing people into low, moderate, high and very high risk, provided an easy clinical interpretation of the predicted risk estimate.

Conclusion

In summary, we presented a prediabetes-specific risk model to identify people at highest risk of CV death. The study was designed as a proof‑of‑concept study to explore the possibility of developing a discriminative model to identify fatal CV events among people with prediabetes and to explore patterns associated with an increased risk. The presented risk model aims to assist clinical decision-making in overweight/obese people with prediabetes having established CVD, thus providing better personalized guidance for early prevention of CV complications in a high-risk population. Time-to-event information and a broader outcome target (including both non-fatal and fatal CV events, such as MACE) could be considered in further analysis to improve clinical utility. Lastly, external validation using real-world data is crucial before considering the prediction model in a real-world clinical setting.

Abbreviations

ADA

American diabetes association

ALT

Alanine aminotransferase

CHF

Chronic heart failure

CV

Cardiovascular

CVD

Cardiovascular disease

eGFR

Estimated glomerular filtration rate

GGT

Gamma glutamyl transferase

hsCRP

High sensitive C-reactive protein

IEC

International expert committee

IFG

Impaired fasting glucose

IGT

Impaired glucose tolerance

MACE

Major adverse cardiovascular events

ROC AUC

Receiver operating characteristic curve area under the curve

Author contributions

AKA, KF, DV, BJS and MHJ conceptualised the study. AKA, SCL and MHJ were responsible for data processing and model development. All authors were responsible for interpretation of the results and data visualisation. All authors reviewed, edited and approved the final manuscript.

Funding

This work was supported by Novo Nordisk A/S and a research grant from the Danish Diabetes and Endocrine Academy, funded by the Novo Nordisk Foundation grant number NNF22SA0079901, and the Danish Data Science Academy, funded by the Novo Nordisk Foundation grant number NNF21SA0069429. Novo Nordisk A/S provided access to the data used in the study. Danish Diabetes and Endocrine Academy and Danish Data Science Academy had no role in the study.

Data availability

The dataset supporting the conclusions of this article will not be made available to others.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

AKA is a Novo Nordisk employee and has received a PhD grant from the DDEA (grant number NNF22SA0079901) and the Danish Data Science Academy (grant number NNF21SA0069429), funded by the Novo Nordisk Foundation. KF is a former Novo Nordisk employee and holds Novo Nordisk shares. DV, BJS, NBJ and MHJ are Novo Nordisk employees and holds Novo Nordisk shares. SLC has received research funding from i-SENS Inc., holds shares in Novo Nordisk, has received consultancy fees from Roche Diagnostics and Medicus Engineering.