Machine learning-based prediction model for predicting the impact of insulin resistance on the risk of ischemic cardiomyopathy Authors: Tuersunjiang Naman, Hui Cheng, Xiao-lin Yu, Zi-tong Guo Published: 2026 ABSTRACT BackgroundThe triglyceride-to-high-density lipoprotein cholesterol (TG/HDL-C) ratio and triglyceride glucose-body mass (TyG-BMI) index are reliable indicators of insulin resistance (IR). This study investigated their association with ischemic cardiomyopathy (ICM) and developed a machine learning-based model for ICM risk prediction.MethodsIn total, 1,603 subjects participated in this study. Univariable logistic regression analysis was conducted, and variables with P < 0.05 were selected for multivariable logistic regression to identify independent risk factors for ICM. Variables meeting this criterion were adopted to create eight machine learning models, from which the optimal model was selected. Using this best-performing model, SHAP values were visualized, and an online calculator was developed. The model was validated via a calibration plot and DCA.ResultsUnivariate and multivariate logistic regression analyses revealed that TyG-BMI, age, ejection fraction, TC/HDL-C, sex, HDL-C, TC, BMI, hemoglobin, diabetes, and hypertension were independent risk factors for ICM (P < 0.05). Based on these factors, SHAP visualization and an online calculator were developed. The calibration plot indicated strong alignment between the model’s predicted and actual values, whereas the DCA demonstrated the model’s clinical utility.ConclusionThe TyG-BMI and TC/HDL-C ratio independently predict ICM risk, with the XGB model identified as the most effective for ICM risk prediction, indicating substantial clinical applicability.Clinical trial registration numberNot applicable.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12911-026-03693-w. Background Ischemic cardiomyopathy (ICM), introduced in 1970 by Burch et al. [1], is a specific or advanced subtype of coronary heart disease that arises from coronary atherosclerotic heart disease [2]. The obstruction or narrowing of coronary arteries impedes myocardial blood flow, preventing it from meeting the metabolic demands of the heart, which typically manifests as chest pain and dyspnea. Prolonged ischemia may result in ongoing myocardial dysfunction or even heart failure. ICM broadly encompasses clinical symptoms related to myocardial ischemia [3]. The prevalence of ischemic heart disease has recently increased sevenfold in urban areas and fourfold in rural areas in India [4], making it a significant contributor to global morbidity and mortality [3]. Identifying risk factors associated with ICM is essential for reducing its incidence. Similar to other complex disorders, ICM arises from interactions among genetic, environmental, and lifestyle factors. These ICM-related factors are diverse and include microvascular dysfunction, inflammation, disruption of Ca2 + homeostasis, and activation of apoptosis [5, 6]. A previous study reported that [7] age, gender, diabetes, hypertension, high-density cholesterol, and low-density high-density cholesterol are risk factors for ICM. Although ICM is recognized as a multifactorial disease with some risk factors already identified, ongoing research continues to explore additional contributing factors. Insulin resistance (IR) is increasingly recognized for its role in the pathophysiology of cardiovascular disease (CVD). Recently, the triglyceride-to-high-density lipoprotein cholesterol (TG/HDL-C) ratio and the triglyceride glucose-body mass (TyG-BMI) index have been shown to be effective and reliable indicators of IR [8], with both metrics showing correlations with certain cardiovascular diseases and diabetes, with the latter being an independent risk factor for ICM. Consequently, the TG/HDL-C ratio and TyG-BMI may also be correlated with ICM risk; however, no prior studies have examined the correlation between TG/HDL-C and TyG-BMI and ICM incidence. Recently, artificial intelligence has been extensively used to identify early prognostic factors for various diseases. Some researchers advocate the application of novel prediction models on the basis of machine learning rather than conventional disease severity classification systems to support more tailored treatment strategies [9–11]. Machine learning leverages algorithms with specific objectives that operate without explicit programming, achieves high performance in the development of predictive models, and is increasingly utilized in medical applications [12]. Although numerous predictive models are available for research, their applicability in clinical practice and effectiveness in risk prediction for disease prognosis remain largely uncertain [13–16], and no systematic tool currently exists to diagnose or predict ICM risk accurately. This study aimed to investigate the correlation of TG/HDL-C and TyG-BMI with ICM risk and develop a novel machine learning-based model for diagnosing and predicting ICM risk, with the goal of supporting clinicians in making personalized clinical decisions. Materials and methods Study design and populations In total, 1659 participants were recruited from People’s Hospital in the Xinjiang Uyghur Autonomous Region from January 2014 to December 2018. Among the 1659 subjects, 1603 individuals were included in the current work, including 904 ICM patients and 699 normal controls, according to the eligibility criteria (Fig. 1). Fig. 1Flowchart of the study Flowchart of the study All the subjects received coronary angiography prior to or during the present hospitalization. ICM patients whose coronary angiography indicated > 50% luminal stenosis in at least one coronary artery, those who previously received percutaneous coronary intervention or coronary artery bypass grafting, those whose HF symptoms were in line with N-terminal pro-B-type natriuretic peptide (NT-pro BNP) at > 125 ng/mL, and those who presented with angina symptoms that might improve following rest or nitroglycerin administration were included. ICM patients were excluded if they had unstable hemodynamics, acute decompensated HF, liver/kidney/autoimmune/hematological disorders, noncardiac disorders with an estimated lifespan < 1 year, cachexia, or were unwilling to participate in the study. To guarantee that the patients were ICM patients, we excluded all kinds of diseases (with the exception of coronary artery disease) that can result in heart failure (NT-pro BNP > 125 ng/mL), such as valvular heart disease, congenital heart disease, renal insufficiency, and hepatic insufficiency. The following normal controls were recruited: those with < 50% coronary artery luminal stenosis verified through coronary angiography by two or more senior cardiologists and those without any additional disorders or angina on exertion. Selection and definition of covariates The covariates included demographic data, clinical manifestations, medical history of the participants, in-hospital test results, and suspected risk factors for ICM. Demographic data, clinical manifestations, and medical history were obtained when patients were admitted to the hospital. Test results were obtained via ultrasound test results and blood test results. Blood was collected from all the subjects on the first day following admission for analysis in the laboratory of the People’s Hospital of Xinjiang Uyghur Autonomous Region. Blood parameters, such as white blood cell (WBC) count, platelet (PLT) count, hemoglobin concentration, blood urea nitrogen (BUN), creatinine (CR), total cholesterol (TC), triglyceride (TG), and high- and low-density lipoprotein-cholesterol (HDL-C and LDL-C), were analyzed. Suspected risk factors for ICM: Smokers were those who had ever smoked within the last 6 months or who had moved for more than 6 months. Drinkers referred to participants who drank alcoholic beverages at 100 g or more every week in the prior month. Hypertension can be diagnosed on the basis of a systolic blood pressure (SBP) ≥ 140 mmHg, a diastolic blood pressure (DBP) ≥ 90 mmHg, and the use of antihypertensive medications over the last 2 weeks, in line with the 2018 European Society of Cardiology (ESC)/European Society of Hypertension (EHS) Guidelines [17]. Diabetes mellitus (DM) was defined as a glucose content ≥ 11.1 mmol/L (200 mg/dL) 2 h after the administration of 75 g of oral glucose or a fasting plasma glucose content ≥ 7.0 mmol/L (126 mg/dL) or the use of antidiabetic medication [18, 19]. Body mass index (BMI) can be determined by dividing weight (kg) by height squared (m2). IR indices can be determined as follows: TyG index = ln [TG (mg/dL)×FBG (mg/dL) ÷ 2]; TG/HDL-C = TG (mg/dL) ÷ HDL-C (mg/dL; TyG-BMI index = TyG×BMI (kg/m2) [20]. Statistical analysis Basic features for all patients in the case and control groups are presented as the means and interquartile ranges, and categorical variables are presented as frequencies and percentages. Chi-square tests (χ2) and Mann-Whitney (Z) tests were adopted for comparing continuous variables and categorical variables, respectively. We performed univariable logistic regression analysis. Then, variables with P < 0.05 were incorporated for multivariate regression to select statistically significant (P < 0.05) variables as independent risk factors for ICM. We randomly classified all the participants into a training group with 1122 (70%) participants and a validation group with 481 participants (30%). We established eight machine learning models, including extreme gradient boosting (XGB), decision tree (DT), random forest (RF), logistic regression (LR), support vector machine (SVM), K-nearest neighbor (KNN), naive Bayes (BNB) and multilayer perceptron (MLP), and evaluated them via 10-fold cross-validation (for the training group) or receiver operating characteristic (ROC) curves, with a higher area under the ROC curve (AUC) representing greater prediction accuracy [21]. SHAP was used to interpret the prediction models, which represent a unified method for calculating the impact or contribution of every feature to the eventual prediction [22]. Moreover, SHAP values reflect the positive or negative contribution of every predictor to the target variable. Additionally, a specific SHAP value set can interpret certain observations within the dataset. We performed model validation in both the training group and the validation group. This study built calibration plots to assess whether the prediction was accurate and whether the predicted values were consistent with real measurements. In the plot, the 45° diagonal line represented great calibration. Calibration plots and Hosmer-Lemeshow tests were adopted for evaluating the calibration performance [23], with a higher P value indicating better model calibration. When the results predicted by the model were very different from the real measurements (P < 0.05), poor calibration was predicted. Clinical utility was evaluated via decision curve analysis (DCA) to determine the net benefits at diverse threshold probabilities [24]. Net benefits were determined through the subtraction of the false-positive patient percentage from the true-positive patient percentage, weighted by the risk of waiving interventions relative to the negative impact of unnecessary interventions. SPSS 27.0, R 4.3.1, and Python 3.10.9 versions were adopted for analyzing the data. Results A total of 1603 subjects were included in this study; 971 (60.57%) participants were male, and 632 (39.43%) were female. We compared the baseline characteristics of all patients in the case and control groups and detected significant differences in age; WBC, RBC, PLT, Hgb, Na+, Cl–, Ca2+, BUN, TG, BMI, TyG-BMI, TC, HDL, TC/HDL, LDL-C, LVED, NTprp-BNP, ejection fraction, sex, smoking status, drinking status, hypertension, and diabetes status (P < 0.05); however, the K+, AST, ALT, Crea, Albumin, total bilirubin, and CRP levels were not significantly different (P > 0.05) (Table 1). Table 1Basic features of ICM or non-ICM participantsVariablesTotal (n = 1603)Control(n = 699)Case(n = 904)Z/χ2 P Age, M (Q1, Q3)52.00 (43.00, 62.00)44.00 (34.00, 51.00)59.00 (51.75, 68.00)–24.22< 0.001WBC, M (Q1, Q3)6.60 (5.60, 7.40)6.30 (5.00, 7.30)6.70 (5.80, 7.53)–5.57< 0.001RBC, M (Q1, Q3)4.70 (4.30, 5.10)4.60 (4.30, 5.00)4.70 (4.40, 5.20)–3.84< 0.001PLT, M (Q1, Q3)229.00 (197.00, 256.50)235.00 (205.00, 255.00)225.00 (189.00, 258.25)–2.720.006Hemoglobin, M (Q1, Q3)136.00 (126.00, 144.00)138.00 (129.00, 144.00)134.00 (121.75, 144.00)–6.16< 0.001Na+, M (Q1, Q3)141.20 (139.30, 143.10)141.30 (139.90, 143.00)141.10 (138.90, 143.10)–2.370.018Cl−, M (Q1, Q3)105.70 (103.60, 107.70)106.00 (103.70, 107.80)105.60 (103.20, 107.70)–2.180.029K+, M (Q1, Q3)3.80 (3.60, 4.10)3.80 (3.50, 4.10)3.90 (3.60, 4.10)–1.740.081Ca2+, M (Q1, Q3)2.20 (2.20, 2.30)2.20 (2.20, 2.30)2.20 (2.20, 2.30)–2.220.027AST, M (Q1, Q3)28.20 (21.90, 44.95)27.90 (22.10, 43.70)28.30 (21.78, 45.40)–0.580.562ALT, M (Q1, Q3)25.50 (18.00, 42.70)25.20 (17.50, 42.70)26.35 (18.30, 42.70)–1.840.066BUN, M (Q1, Q3)5.90 (4.90, 7.10)5.70 (4.70, 6.90)6.00 (4.90, 7.30)–2.620.009Cr, M (Q1, Q3)67.60 (58.80, 81.80)68.50 (58.40, 82.00)67.30 (59.20, 81.25)–0.100.922Albumin, M (Q1, Q3)40.40 (38.60, 43.20)39.70 (38.60, 42.50)40.90 (38.20, 43.50)–1.840.066Total bilirubin, M (Q1, Q3)11.90 (8.60, 15.40)11.70 (8.60, 14.80)12.20 (8.50, 16.30)–1.910.056TG, M (Q1, Q3)1.20 (1.10, 1.80)1.10 (1.00, 1.30)1.50 (1.10, 3.10)–14.14< 0.001BMI, M (Q1, Q3)28.00 (24.20, 31.00)27.00 (24.05, 31.00)28.00 (24.78, 31.25)–3.060.002TyG-BMI, M (Q1, Q3)37.56 (24.68, 70.72)25.32 (18.53, 31.64)63.61 (42.17, 87.15)–28.25< 0.001TC, M (Q1, Q3)3.70 (2.80, 4.80)3.50 (2.70, 4.20)4.00 (2.90, 6.50)–9.34< 0.001HDL-C, M (Q1, Q3)1.10 (0.90, 1.20)1.10 (1.00, 1.30)1.00 (0.80, 1.20)–9.24< 0.001TG/HDL-C, M (Q1, Q3)3.57 (2.71, 4.86)3.13 (2.51, 3.91)4.25 (2.90, 5.91)–13.28< 0.001LDL-C, M (Q1, Q3)2.40 (1.90, 3.00)2.40 (1.90, 2.90)2.50 (1.90, 3.10)–3.49< 0.001CRP, M (Q1, Q3)5.10 (4.80, 6.50)5.10 (5.00, 5.80)5.10 (2.48, 7.10)–0.580.564LVDD, M (Q1, Q3)50.00 (47.00, 54.00)48.00 (46.00, 50.00)52.00 (48.00, 61.00)–16.12< 0.001Ntpro-BNP, M (Q1, Q3)141.00 (100.00, 193.50)131.00 (65.00, 183.00)145.00 (129.00, 214.32)–7.22< 0.001Ejection Fraction, M (Q1, Q3)61.90 (54.60, 63.80)62.60 (61.40, 64.15)59.65 (41.77, 63.40)–13.36< 0.001Sex, n(%)269.93< 0.001 Male971 (60.57)264 (37.77)707 (78.21) Female632 (39.43)435 (62.23)197 (21.79)Smoking, n(%)53.07< 0.001 No1242 (77.48)602 (86.12)640 (70.80) Yes361 (22.52)97 (13.88)264 (29.20)Drinking, n(%)19.60< 0.001 No1425 (88.90)649 (92.85)776 (85.84) Yes178 (11.10)50 (7.15)128 (14.16)Hypertension, n(%)77.04< 0.001 No650 (40.55)369 (52.79)281 (31.08) Yes953 (59.45)330 (47.21)623 (68.92)Diabet, n(%)84.72< 0.001 No1245 (77.67)619 (88.56)626 (69.25) Yes358 (22.33)80 (11.44)278 (30.75) Table 1Basic features of ICM or non-ICM participantsVariablesTotal (n = 1603)Control(n = 699)Case(n = 904)Z/χ2 P Age, M (Q1, Q3)52.00 (43.00, 62.00)44.00 (34.00, 51.00)59.00 (51.75, 68.00)–24.22< 0.001WBC, M (Q1, Q3)6.60 (5.60, 7.40)6.30 (5.00, 7.30)6.70 (5.80, 7.53)–5.57< 0.001RBC, M (Q1, Q3)4.70 (4.30, 5.10)4.60 (4.30, 5.00)4.70 (4.40, 5.20)–3.84< 0.001PLT, M (Q1, Q3)229.00 (197.00, 256.50)235.00 (205.00, 255.00)225.00 (189.00, 258.25)–2.720.006Hemoglobin, M (Q1, Q3)136.00 (126.00, 144.00)138.00 (129.00, 144.00)134.00 (121.75, 144.00)–6.16< 0.001Na+, M (Q1, Q3)141.20 (139.30, 143.10)141.30 (139.90, 143.00)141.10 (138.90, 143.10)–2.370.018Cl−, M (Q1, Q3)105.70 (103.60, 107.70)106.00 (103.70, 107.80)105.60 (103.20, 107.70)–2.180.029K+, M (Q1, Q3)3.80 (3.60, 4.10)3.80 (3.50, 4.10)3.90 (3.60, 4.10)–1.740.081Ca2+, M (Q1, Q3)2.20 (2.20, 2.30)2.20 (2.20, 2.30)2.20 (2.20, 2.30)–2.220.027AST, M (Q1, Q3)28.20 (21.90, 44.95)27.90 (22.10, 43.70)28.30 (21.78, 45.40)–0.580.562ALT, M (Q1, Q3)25.50 (18.00, 42.70)25.20 (17.50, 42.70)26.35 (18.30, 42.70)–1.840.066BUN, M (Q1, Q3)5.90 (4.90, 7.10)5.70 (4.70, 6.90)6.00 (4.90, 7.30)–2.620.009Cr, M (Q1, Q3)67.60 (58.80, 81.80)68.50 (58.40, 82.00)67.30 (59.20, 81.25)–0.100.922Albumin, M (Q1, Q3)40.40 (38.60, 43.20)39.70 (38.60, 42.50)40.90 (38.20, 43.50)–1.840.066Total bilirubin, M (Q1, Q3)11.90 (8.60, 15.40)11.70 (8.60, 14.80)12.20 (8.50, 16.30)–1.910.056TG, M (Q1, Q3)1.20 (1.10, 1.80)1.10 (1.00, 1.30)1.50 (1.10, 3.10)–14.14< 0.001BMI, M (Q1, Q3)28.00 (24.20, 31.00)27.00 (24.05, 31.00)28.00 (24.78, 31.25)–3.060.002TyG-BMI, M (Q1, Q3)37.56 (24.68, 70.72)25.32 (18.53, 31.64)63.61 (42.17, 87.15)–28.25< 0.001TC, M (Q1, Q3)3.70 (2.80, 4.80)3.50 (2.70, 4.20)4.00 (2.90, 6.50)–9.34< 0.001HDL-C, M (Q1, Q3)1.10 (0.90, 1.20)1.10 (1.00, 1.30)1.00 (0.80, 1.20)–9.24< 0.001TG/HDL-C, M (Q1, Q3)3.57 (2.71, 4.86)3.13 (2.51, 3.91)4.25 (2.90, 5.91)–13.28< 0.001LDL-C, M (Q1, Q3)2.40 (1.90, 3.00)2.40 (1.90, 2.90)2.50 (1.90, 3.10)–3.49< 0.001CRP, M (Q1, Q3)5.10 (4.80, 6.50)5.10 (5.00, 5.80)5.10 (2.48, 7.10)–0.580.564LVDD, M (Q1, Q3)50.00 (47.00, 54.00)48.00 (46.00, 50.00)52.00 (48.00, 61.00)–16.12< 0.001Ntpro-BNP, M (Q1, Q3)141.00 (100.00, 193.50)131.00 (65.00, 183.00)145.00 (129.00, 214.32)–7.22< 0.001Ejection Fraction, M (Q1, Q3)61.90 (54.60, 63.80)62.60 (61.40, 64.15)59.65 (41.77, 63.40)–13.36< 0.001Sex, n(%)269.93< 0.001 Male971 (60.57)264 (37.77)707 (78.21) Female632 (39.43)435 (62.23)197 (21.79)Smoking, n(%)53.07< 0.001 No1242 (77.48)602 (86.12)640 (70.80) Yes361 (22.52)97 (13.88)264 (29.20)Drinking, n(%)19.60< 0.001 No1425 (88.90)649 (92.85)776 (85.84) Yes178 (11.10)50 (7.15)128 (14.16)Hypertension, n(%)77.04< 0.001 No650 (40.55)369 (52.79)281 (31.08) Yes953 (59.45)330 (47.21)623 (68.92)Diabet, n(%)84.72< 0.001 No1245 (77.67)619 (88.56)626 (69.25) Yes358 (22.33)80 (11.44)278 (30.75) Basic features of ICM or non-ICM participants We further performed univariate logistic regression analysis on all the participants and detected significant differences in sex, smoking status, drinking status, hypertension status, diabetes status, age, WBC count, Hgb, Na+, Cl+, Ca2+, PLT, AST, ALT, total bilirubin, TG, TyG-BMI, BMI, HDL-C, TC, LDL-C, TC/HDL-C, Ntpro-BNP, LVED, and ejection fraction between the case and control groups (P < 0.05), and RBC, K+, Crea, bilirubin, BUN, and CRP were not significantly different (P > 0.05). We selected variables that were significantly different between the two groups and performed multivariable logistic regression analysis. The results revealed that sex, hypertension, diabetes status, age, Hgb, TyG-BMI, BMI, HDL, TC, TC/HDL, and ejection fraction were independent risk factors for ICM (P < 0.05) (Table 2). Table 2Univariable and multivariable logistic regression for ICM-related risk factorsVariablesUnivariate logisticMultivariate logisticβS.EZ P OR (95%CI)βS.EZ P OR (95%CI)Sex MaleReferenceReference Female-1.780.11-15.85< 0.0010.17 (0.14 ~ 0.21)-1.510.30-5.10< 0.0010.22 (0.12 ~ 0.40)Smoking NoReferenceReference Yes0.940.137.14< 0.0012.56 (1.98 ~ 3.31)0.190.380.490.6251.21 (0.57 ~ 2.56)Drinking NoReferenceReference Yes0.760.184.35< 0.0012.14 (1.52 ~ 3.02)0.220.480.460.6431.25 (0.49 ~ 3.21)Hypertension NoReferenceReference Yes0.910.108.69< 0.0012.48 (2.02 ~ 3.04)0.950.253.82< 0.0012.58 (1.59 ~ 4.21)Diabet NoReferenceReference Yes1.230.148.88< 0.0013.44 (2.62 ~ 4.51)0.720.332.180.0292.06 (1.07 ~ 3.96)Age0.130.0119.52< 0.0011.14 (1.12 ~ 1.15)0.140.0111.46< 0.0011.15 (1.12 ~ 1.18)RBC0.080.051.630.1031.08 (0.98 ~ 1.19)WBC0.140.034.63< 0.0011.15 (1.09 ~ 1.23)0.010.080.120.9051.01 (0.87 ~ 1.17)Hemoglobin-0.020.00-6.73< 0.0010.98 (0.97 ~ 0.98)-0.040.01-4.43< 0.0010.96 (0.94 ~ 0.98)Na+-0.030.01-2.450.0140.97 (0.94 ~ 0.99)-0.000.04-0.060.9541.00 (0.93 ~ 1.08)Cl–-0.050.02-3.130.0020.95 (0.93 ~ 0.98)0.100.052.270.1101.10 (0.98 ~ 1.22)K+0.040.110.400.6891.05 (0.84 ~ 1.30)Ca2+0.500.252.040.0411.65 (1.02 ~ 2.67)-0.060.63-0.090.9300.95 (0.28 ~ 3.23)PLT-0.010.00-2.370.0180.99 (0.99 ~ 0.99)-0.000.00-1.370.1721.00 (0.99 ~ 1.00)AST0.010.002.580.0101.01 (1.01 ~ 1.01)0.010.011.420.1551.01 (1.00 ~ 1.02)ALT0.010.002.520.0121.01 (1.01 ~ 1.01)-0.000.01-0.510.6101.00 (0.99 ~ 1.01)Crea0.000.000.880.3811.00 (1.00 ~ 1.01)Albumin0.010.010.550.5791.01 (0.98 ~ 1.04)Total bilirubin0.030.014.10< 0.0011.03 (1.02 ~ 1.05)0.000.020.260.7951.00 (0.97 ~ 1.04)BUN0.000.020.060.9521.00 (0.97 ~ 1.03)TG1.440.1211.76< 0.0014.22 (3.32 ~ 5.36)-0.040.22-0.200.8420.96 (0.62 ~ 1.47)TyG-BMI0.120.0118.41< 0.0011.13 (1.11 ~ 1.14)0.120.019.89< 0.0011.13 (1.10 ~ 1.16)BMI0.020.012.280.0231.02 (1.01 ~ 1.04)0.010.011.540.0431.02 (1.01 ~ 1.03)HDL-C-1.570.20-7.88< 0.0010.21 (0.14 ~ 0.31)-0.820.41-1.990.0460.44 (0.20 ~ 0.99)TG0.450.0411.57< 0.0011.57 (1.45 ~ 1.69)4.061.382.940.00358.24 (3.88 ~ 873.32)LDL-C0.420.066.85< 0.0011.52 (1.35 ~ 1.72)0.010.160.070.9451.01 (0.73 ~ 1.39)TC/HDL-C0.530.0413.49< 0.0011.70 (1.58 ~ 1.84)1.440.453.230.0014.24 (1.76 ~ 10.20)CRP-0.000.01-0.100.9171.00 (0.99 ~ 1.01)NT-proBNP0.010.006.17< 0.0011.01 (1.01 ~ 1.01)0.000.000.760.4501.00 (1.00 ~ 1.00)LVDD0.170.0114.54< 0.0011.18 (1.16 ~ 1.21)0.060.031.960.0511.06 (1.01 ~ 1.12)Ejection Fraction-0.130.01-14.03< 0.0010.88 (0.86 ~ 0.89)-0.090.02-4.27< 0.0010.91 (0.87 ~ 0.95) Table 2Univariable and multivariable logistic regression for ICM-related risk factorsVariablesUnivariate logisticMultivariate logisticβS.EZ P OR (95%CI)βS.EZ P OR (95%CI)Sex MaleReferenceReference Female-1.780.11-15.85< 0.0010.17 (0.14 ~ 0.21)-1.510.30-5.10< 0.0010.22 (0.12 ~ 0.40)Smoking NoReferenceReference Yes0.940.137.14< 0.0012.56 (1.98 ~ 3.31)0.190.380.490.6251.21 (0.57 ~ 2.56)Drinking NoReferenceReference Yes0.760.184.35< 0.0012.14 (1.52 ~ 3.02)0.220.480.460.6431.25 (0.49 ~ 3.21)Hypertension NoReferenceReference Yes0.910.108.69< 0.0012.48 (2.02 ~ 3.04)0.950.253.82< 0.0012.58 (1.59 ~ 4.21)Diabet NoReferenceReference Yes1.230.148.88< 0.0013.44 (2.62 ~ 4.51)0.720.332.180.0292.06 (1.07 ~ 3.96)Age0.130.0119.52< 0.0011.14 (1.12 ~ 1.15)0.140.0111.46< 0.0011.15 (1.12 ~ 1.18)RBC0.080.051.630.1031.08 (0.98 ~ 1.19)WBC0.140.034.63< 0.0011.15 (1.09 ~ 1.23)0.010.080.120.9051.01 (0.87 ~ 1.17)Hemoglobin-0.020.00-6.73< 0.0010.98 (0.97 ~ 0.98)-0.040.01-4.43< 0.0010.96 (0.94 ~ 0.98)Na+-0.030.01-2.450.0140.97 (0.94 ~ 0.99)-0.000.04-0.060.9541.00 (0.93 ~ 1.08)Cl–-0.050.02-3.130.0020.95 (0.93 ~ 0.98)0.100.052.270.1101.10 (0.98 ~ 1.22)K+0.040.110.400.6891.05 (0.84 ~ 1.30)Ca2+0.500.252.040.0411.65 (1.02 ~ 2.67)-0.060.63-0.090.9300.95 (0.28 ~ 3.23)PLT-0.010.00-2.370.0180.99 (0.99 ~ 0.99)-0.000.00-1.370.1721.00 (0.99 ~ 1.00)AST0.010.002.580.0101.01 (1.01 ~ 1.01)0.010.011.420.1551.01 (1.00 ~ 1.02)ALT0.010.002.520.0121.01 (1.01 ~ 1.01)-0.000.01-0.510.6101.00 (0.99 ~ 1.01)Crea0.000.000.880.3811.00 (1.00 ~ 1.01)Albumin0.010.010.550.5791.01 (0.98 ~ 1.04)Total bilirubin0.030.014.10< 0.0011.03 (1.02 ~ 1.05)0.000.020.260.7951.00 (0.97 ~ 1.04)BUN0.000.020.060.9521.00 (0.97 ~ 1.03)TG1.440.1211.76< 0.0014.22 (3.32 ~ 5.36)-0.040.22-0.200.8420.96 (0.62 ~ 1.47)TyG-BMI0.120.0118.41< 0.0011.13 (1.11 ~ 1.14)0.120.019.89< 0.0011.13 (1.10 ~ 1.16)BMI0.020.012.280.0231.02 (1.01 ~ 1.04)0.010.011.540.0431.02 (1.01 ~ 1.03)HDL-C-1.570.20-7.88< 0.0010.21 (0.14 ~ 0.31)-0.820.41-1.990.0460.44 (0.20 ~ 0.99)TG0.450.0411.57< 0.0011.57 (1.45 ~ 1.69)4.061.382.940.00358.24 (3.88 ~ 873.32)LDL-C0.420.066.85< 0.0011.52 (1.35 ~ 1.72)0.010.160.070.9451.01 (0.73 ~ 1.39)TC/HDL-C0.530.0413.49< 0.0011.70 (1.58 ~ 1.84)1.440.453.230.0014.24 (1.76 ~ 10.20)CRP-0.000.01-0.100.9171.00 (0.99 ~ 1.01)NT-proBNP0.010.006.17< 0.0011.01 (1.01 ~ 1.01)0.000.000.760.4501.00 (1.00 ~ 1.00)LVDD0.170.0114.54< 0.0011.18 (1.16 ~ 1.21)0.060.031.960.0511.06 (1.01 ~ 1.12)Ejection Fraction-0.130.01-14.03< 0.0010.88 (0.86 ~ 0.89)-0.090.02-4.27< 0.0010.91 (0.87 ~ 0.95) Univariable and multivariable logistic regression for ICM-related risk factors We randomly classified all the participants into a training group with 1122 (70%) participants and a validation group with 481 (30%) participants and compared the basic features of the subjects from both the training and validation groups. There were no significant differences in any of the variables (Supplementary Table 1). Establishing and evaluating the model We used variables that could independently predict ICM risk to establish eight machine learning models and verified them via a 10-fold cross-validation test in the training set. The results indicated that the XGB model had the highest AUC value (0.98) among all the models (Fig. 2). Fig. 210-fold cross-validation of 8 machine learning models 10-fold cross-validation of 8 machine learning models We generated ROC curves for all the models from the training group and validation group. The results indicated that the XGB model had the highest AUC value in the training group (0.98) and validation group (0.96) (Fig. 3). Fig. 3ROC curves of eight machine learning models for the training (A) and validation sets (B) ROC curves of eight machine learning models for the training (A) and validation sets (B) The BNB model also has the same AUC value (0.96) as the XGB model in the validation set. After we integrated the results of 10-fold cross-validation and the ROC curves of the eight models for both groups, we determined that the XGB model was one of the more appropriate models in this study. Thus, the XGB model was used to develop the SHAP model (Fig. 4). Fig. 4SHAP values for the XGBoost model output. Every point represents one variable for each observation. The point color is related to the corresponding relative height within this variable. Blue and red indicate low and high levels respectively SHAP values for the XGBoost model output. Every point represents one variable for each observation. The point color is related to the corresponding relative height within this variable. Blue and red indicate low and high levels respectively Shap presents the independent risk factors for ICM according to its degree of importance (from TyG-BMI to hypertension). We also show how every feature contributes to the final prediction, which can be employed to explain the model predictions of different patients. Moreover, we provide two typical examples to illustrate the interpretability of the model: An individual in the case group had TG/HDL-C = 4.33 and TyG-BMI = 22.99, and another individual in the case group had TG/HDL-C = 2.88 and TyG-BMI = 18.204 (Fig. 5A and B). The blue and red colors indicate whether the factor reduced (blue) or increased (red) the risk of ICM. The combined effects of all factors provided the final SHAP prediction probability; for two representative individuals, there was a high prediction probability of 0.94 and a low prediction probability of 0.31. Fig. 5Interpretation of the model prediction results for two typical samples: (A) An individual with a high risk (0.94) of ICM. (B) An individual who has a low risk (0.31) of ICM Interpretation of the model prediction results for two typical samples: (A) An individual with a high risk (0.94) of ICM. (B) An individual who has a low risk (0.31) of ICM Model validation We validated the model by evaluating the DCA and calibration plot. The calibration plot of the model revealed high consistency of the model-predicted values with the real measurements in the two groups (Fig. 6A, B). Fig. 6Calibration plot of eight machine learning models. (A) In the training set. (B) In the validation set Calibration plot of eight machine learning models. (A) In the training set. (B) In the validation set In accordance with the DCA, if the constructed nomogram was used to predict ICM risk, a greater net benefit could be observed than with the “treat all” or “treat none” strategies at the 0.0–1.0 threshold probabilities, thus indicating that the developed model has greater clinical utility (Fig. 7A, B). Fig. 7DCA plots of eight machine learning models. (A) In the training set. (B) In the validation set DCA plots of eight machine learning models. (A) In the training set. (B) In the validation set Discussion In this single-center retrospective study, we found that TyG-BMI, age, ejection fraction, TC/HDL-C, sex, HDL-C, TC, BMI, hemoglobin, diabetes, and hypertension were independent risk factors for ICM. The TyG-BMI and TC/HDL-C other indices are established risk factors for ICM; as we know, the present work is the first to reveal the correlation of TyG-BMI and TC/HDL-C with ICM. Every one-unit increase in the TyG-BMI results in an increase in ICM risk of 0.13 times, and with an increase in the TC/HDL-C ratio, the risk of ICM increases 3.24 times. ICM is the heart failure stage of coronary artery disease, and it is the most common type of CVD [25]. In accordance with a previous study, high IR increases the risk of CVD and directly participates in the pathophysiological process of CVD [8]. A higher IR represents a higher TG/HDL-C ratio, a higher TyG-BMI and a higher TyG. These indices also increase the risk of CVDs [26, 27], but no previous study has specifically clarified their relationships with ICM risk. This work is the first to report the correlations of the TG/HDL-C ratio and TyG-BMI with ICM risk. TyG, an independent risk factor for cardiovascular disease, is also directly related to the TyG-BMI. An individual who has a higher TyG has a greater risk of diabetes. According to the results of this research, both a high TyG-BMI and diabetes are independent risk factors for ICM; thus, the results of this research are highly consistent with those of previous studies. The underlying mechanisms by which high IR can increase the risk of ICM are as follows: high IR can decrease or result in decreased insulin sensitivity in target tissues and organs, which is manifested by impaired glucose absorption and oxidization [28, 29]; high IR leads to endothelial impairment [28], ectopic angiotensinogen generation [30], or even renin-angiotensin-aldosterone system (RAAS) activation [31] and systemic low-grade inflammation, which exacerbates HF progression [32, 33]. The IR state can result in excessive accumulation of lipids within cardiomyocytes, causing a phenomenon called “cardiac lipotoxicity”, which can thus result in cellular dysfunction, myocardial metabolic impairment, and cardiomyocyte apoptosis. These alterations can induce structural and functional changes in cardiomyocytes [34]. On the other hand, IR may decrease glucose bioavailability while inducing alterations in substrate metabolism, resulting in greater myocardial oxygen consumption and lower myocardial compensatory ability [35–37]. The above physiological roles of high IR may increase the risk of ICM accumulation. Here, there is a need to control the IR at an appropriate level to reduce the risk of ICM, and these IR indices are useful variables for controlling ICM risk as well as early prediction of ICM. Artificial intelligence results, including those of machine learning technologies, are capable of managing many variables interactively and nonlinearly [38]. As a game theory approach, SHAP (SHapley additive exPlanations) contributes to the intuitive and accurate interpretation of machine learning model output [39], which reflects the contribution of every feature to pushing model output from the base value. It is the quantified contribution [39]. The feature contribution and the feature with the highest contribution can be found. The features are arranged according to their importance determined through the average contribution of one observation (consistent with that determined by the conventional approach) [40]. In the field of medicine, machine learning technologies are interpreted according to their clinical utility. Variables were chosen on the basis of SHAP feature importance in line with the Shapley value, which has good theoretical properties and fast computation [38]. In our study, we also performed eight machine learning models and selected the best model (XGB model) to diagnose and predict the ICM. This model has good discriminative ability among all the models. The AUCs of these models in the training set were 0.98 and 0.96 in the validation set, and we performed model validation by performing a calibration plot and DCA. As revealed by the calibration plot of the model, the values predicted by this model were highly consistent with the real measurements in the training and validation sets. The DCA results indicated that this model has good clinical usefulness in the training and validation groups. This research has the following limitations. First, the present one-arm retrospective study had a small sample size. Second, this study was not validated externally; thus, future research needs to validate the results externally. In conclusion, the TG/HDL-C ratio and TyG-BMI are related to ICM risk, and the XGB model is the best model with potentially important implications for predicting ICM risk. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1: Supplementary Table 1 Comparison of the baseline characteristics of the training group and the validation group Supplementary Material 1: Supplementary Table 1 Comparison of the baseline characteristics of the training group and the validation group