Between January 2022 to September 2024, We retrospectively reviewed T2DM patients with myocardial ischemia identified through SPECT at our hospital. The criteria for exclusion were listed below: Patients previously treated with percutaneous coronary intervention, coronary bypass surgery, and cerebrovascular diseases before SPECT, SPECT image quality was inadequate for ischemia diagnosis; deficient clinical data; Patients with heart disease, respiratory failure, severe liver and kidney diseases, cancer, severe infections, and other illnesses. Finally, 637 patients with T2DM were incorporated into the study.

All patients underwent gated resting SPECT MPI using a single IQ-SPECT dual-probe scanner (Symbia T16, Siemens, Germany). After fasting for at least 4 hours, 99mTc-MIBI (740–925 MBq) was intravenously injected. Fatty meals were consumed 15 to 20 minutes prior to imaging. Electrodes were placed on the chest to capture gated data. Patients remained supine with arms raised, and acquisition began 60 minutes post-injection, lasting 8 minutes. Imaging parameters included a 128×128 matrix, 208° rotation, 34 frames, and 25 seconds per frame. Images were reconstructed using ordered subsets expectation maximization in fully automated mode.

The SPECT and gated SPECT images were anonymized and visually analyzed by two experienced observers. Myocardial perfusion was assessed by consensus using a 17-segment division of the left ventricle (20) and a four-point grading system: 0, normal uptake; 1, equivocal; 2, moderate reduction; and 3, severe reduction. Segments with an uptake score ≥2 at stress were classified as having a definite uptake reduction (21–23). Only segments with an uptake score ≥2 at stress were considered to have a definite uptake reduction at stress. Segments with an uptake score ≥2 at stress were classified as having a definite uptake reduction. Total myocardial counts were determined in each acquisition with a manually adjusted elliptical region of interest ( Figure 1 ).

The elbow method combined with the Silhouette coefficient were utilized to cluster the 637 T2DM patients based on their clinical characteristics ( Figure 2 ), to identify the appropriate cluster count. The core principle of this method is to reduce to the smallest possible value the sum of squared errors between the cluster center and the points within each cluster. As K grows, the model’s separation becomes more distinct. However, when K exceeds the best appraisal, The cumulative squared errors no longer shows significant changes. K-means is a clustering technique used for grouping objects that consist of both quantitative and qualitative data (24).

The fundamental procedures of the Silhouette coefficient and the elbow method are described below:

(1) Select a Range of K values.

Determine a possible range for K, typically starting from 2(since the Silhouette coefficient is meaningless for K = 1)up to a reasonable maximum value.

(2) Perform the K-means clustering.

For each K value, run the K-means algorithm and record the following results:

SSE(Sum of squared Errors):Used for the Elbow Method.

Silhouette Coefficient: Used to evaluated clustering quality.

(3) Analyze the Elbow Method: Calculate SSE, Plot the Elbow Point, Identify the Elbow point.

(4) Analyze the Silhouette Coefficient: Calculate the Silhouette Coefficient, Plot the Silhouette curve, Select The Optimal K Value.

(5) Compare the Elbow Method and Silhouette Coefficient.

To determine the optimal number of clusters, we evaluated both the elbow method and the silhouette coefficient across a range of K values (K = 2–6). The elbow plot showed that the reduction in the sum of squared errors plateaued after K = 2, while the silhouette coefficient reached its maximum at K = 2. These complementary results indicated that two clusters provided the best balance between separation and stability. Moreover, the two-cluster solution yielded clinically meaningful subgroups with distinct demographic, biochemical, and imaging characteristics, further validating the selection of K = 2.

By using the unsupervised clustering method, the optimal number of clusters that best explained the overall variance in the data was determined.(cluster group 1: n=262, cluster group 2: n=375). The clinical variables and myocardial ischemia characteristics were analyzed and compared across the two cluster subgroups. The analysis was based on Python 3.9.12 with the following libraries: scikit-learn 1.0.2 for implementing K-means clustering and calculating the Silhouette Coefficient, NumPy 1.21.5 for numerical computations, and Matplotlib 3.5.1 for visualizing the Elbow Method and Silhouette Score plots. The Jupyter Notebook environment was used for code development and execution ( Figure 3 ).

Following cluster group identification, clinical information and myocardial ischemia characteristics were compared between the groups. Statistical analyses were conducted with SPSS software (version 25.0; IBM, Armonk, NY, USA). Categorical variables are presented as counts (%), while continuous variables are expressed as mean ± standard deviation. T-tests and chi-square tests were used to compare clinical and SPECT characteristics between cluster groups. A two-tailed P value < 0.05 was considered statistically significant.

The study included a total of 637 individuals with T2DM,of whom 68.3% (435/637) were men, with an average age of 61.65 ± 14.05 years old. Unsupervised K-means clustering analysis was performed on 29 clinical parameters of 637 T2DM subjects. The silhouette coefficient for different values of K was calculated, and the K value with the highest silhouette coefficient was selected as the optimal number of clusters. The results revealed the existence of two clinical subtypes of T2DM with myocardial ischemia, which were classified into Cluster 1 (n=262) and Cluster 2 (n=375). Table 1 presents the clinical characteristics of the participants in the two cluster groups.

The characteristics of myocardial ischemia are compared across the two clustering groups in Table 2 and Figure 4 . In terms of Ischemia areas, cluster group 1 had the higher proportion of patients with incomplete ischemia (69.8%), Cluster group 2 had higher proportion of complete ischemia (11.5%) ( Figure 4A ). Wall motion in cluster group 1 tend to had higher proportion of Hypokinesis and Akinesis, while normal wall motion in cluster group 2 seems more common ( Figure 4C ). For Ventricular contraction, cluster group 1 had worse contraction ( Figure 4D ). In terms of Cardiac systolic coordination, group 1 had a worse proportion ( Figure 4E ). From the patient perspective, cluster group 1 exhibited a greater rate of abnormal ESV and LVEF ( Figure 4F ). No significant variations were noted in the involved segments, EDV, ischemia areas, and cavity size ( Figure 4B ) between the two cluster groups (all P values>0.05).

It is widely recognized that diabetic patients are at greater risk for more aggressive vascular disease, including diffuse coronary atherosclerosis, and exhibit a significantly increased occurrence of heart failure, myocardial infarction (MI), and cardiovascular mortality (27, 28). Diabetic patients have multitude of characteristic features, the interaction of these contributor on myocardial ischemia are attracting more focus. Preceding studies primarily aimed at the limited factors for myocardial ischemia (29–31). The target of unsupervised machine learning is to uncover clusters of patients with analogous combinations of features, free from biases introduced by clinical experts or information on future outcomes. As clinical data continue to grow rapidly, clustering methods may become increasingly valuable for analyzing the varied and multifaceted data available processing the diverse and heterogeneous data found in digital health data.

The clustering approach demonstrated the ability to not only differentiate T2DM patients with diverse clinical profiles but also to indirectly identify distinct subgroups exhibiting various types of myocardial ischemia. The findings revealed that cluster 2 had higher proportion of complete ischemia. This can be attributed to the evidence that cluster group 2 consisted of a larger number of males, exhibited more unhealthy habits such as nicotine use and alcohol consumption, and had elevated levels of LDL-C. Present smoking has been classified as a threat for myocardial ischemia (32). Mild alcohol consumption is known to be cardio protective compared with either heavy drinking or complete abstinence (33, 34). Treatment to reduce LDL cholesterol HDL-C levels is beneficial to improve ischemia (35). This result draws attention to the crucial role of effective management for T2DM patients exhibiting these risk factors for myocardial ischemia.

In line with these risk profiles, patients in Cluster 2 were more frequently prescribed GLP-1 receptor agonists and statins. GLP-1 receptor agonists have been reported to improve coronary microvascular function, while statins effectively reduce LDL-C concentrations. Nevertheless, in this cohort statin therapy was primarily administered at low-to-moderate intensities, which may have attenuated their cardioprotective effect. This limitation could partly explain the persistently higher prevalence of complete ischemia in Cluster 2, despite the seemingly more optimized pharmacological regimen.

Group1 had a higher proportion of hypokinesis and akinesis wall motion, and ventricular contraction and cardiac systolic coordination also worse in group 1. This might be attributed to the fact that patients in Cluster 1 were older (63.23 ± 12.31 vs60.49 ± 15.04 years), had a longer duration of diabetes (10.27 ± 8.77vs5.51 ± 7.61 years), and exhibited poorer glycemic control, as reflected by higher HbA1c levels (7.69 ± 1.76vs6.35 ± 1.68).A previous study showed that wall motion is an independent predictor of ischemic heart (36, 37). In both cross-sectional and short-term longitudinal studies involving older adults, the status of glycemic dysregulation add to the risk of wall motion (38, 39). Additionally, inadequate control of diabetes mellitus (DM) was associated with subclinical left ventricular (LV) dysfunction (40). Traditionally, aging is regarded as a risk factor for myocardial ischemia. We infer from our data that longstanding glycemic abnormality produces a compounded harmful influence on LV wall motion.

Although SGLT2 inhibitors were more commonly prescribed in Cluster 1 as a cardioprotective strategy, the prevalence of reduced exercise tolerance and exercise incapacity remained high. This apparent inconsistency highlights that, despite the demonstrated efficacy of SGLT2 inhibitors in randomized controlled trials, real-world effectiveness may be attenuated due to suboptimal adherence, heterogeneity in therapeutic responsiveness, or patient-specific factors. These findings underscore the need for further evaluation of treatment strategies in routine clinical practice to mitigate LV dysfunction in high-risk T2DM populations.

Although some differences were observed in ventricular contraction, cardiac systolic coordination, ESV and LVEF, there was no statistically significant difference in the ischemic areas between the two cluster groups. A previous study demonstrated that T2DM increase the risk of death among patients with ischemic heart disease. Another studies showed that risk factors for myocardial ischemia, such as hypertension, hyperlipidemia, diabetes, smoking, obesity, age, gender, family history, can exacerbate the severity of myocardial ischemia through different mechanisms (41–43). These risk factors collectively contribute to the increased severity of myocardial ischemia. Therefore, early intervention and comprehensive management targeting these factors are crucial.

Beyond systolic function, diastolic impairment also represents a critical concern in T2DM. The disease promotes myocardial fibrosis and increases ventricular stiffness, thereby contributing to diastolic dysfunction—a hallmark mechanism of heart failure with preserved ejection fraction (HFpEF). In line with this, Cluster 1 patients demonstrated more pronounced abnormalities in left ventricular end-diastolic volume (EDV) and E/E′ ratio, indicating a heightened susceptibility to diastolic dysfunction. Importantly, these patients also exhibited higher mean systolic blood pressure, longer disease duration, and poorer glycemic control, whereas Cluster 2 patients presented with slightly lower diastolic pressure. Consequently, a wider pulse pressure was evident in Cluster 1. Given that both hypertension and diabetes are key risk factors for HFpEF, these pathophysiological distinctions provide additional mechanistic support for the clustering results.

The present study demonstrates the feasibility of applying unsupervised clustering to classify T2DM patients with myocardial ischemia and underscores its potential clinical relevance. By identifying subgroups with distinct ischemic and metabolic features, clustering provides an evidence-based approach for personalized decision-making, enabling physicians to tailor therapies to specific risk profiles. Its integration of SPECT-derived functional parameters further highlights potential for early ischemia detection, thereby improving diagnostic accuracy and guiding timely interventions in high-risk populations. Beyond diagnosis, clustering may assist in predicting disease progression and stratifying cardiovascular risk, offering a valuable tool for long-term management and proactive prevention. Overall, these findings suggest that unsupervised learning could complement conventional risk assessment and foster a more precise and individualized model of care for patients with T2DM.