In this analysis, our principal findings are as follows: (1) HF in patients with DM can be effectively identified with the use of a small subset of the most discriminative parameters exploited to build a MLR model; (2) in general population of patients with DM regardless of their MASLD status, utilizing 5 clinical parameters which are easy-to-obtain in clinical practice (age, type of DM, AF, hyperuricemia, and eGFR) was enough to identify patients who have concomitant HF; (3) in patients with DM and comorbid MASLD, using only three parameters (AF, hyperuricemia, and eGFR) was enough to identify patients who present with HF; and (4) in patients with DM without comorbid MASLD, using only two clinical parameters (age and eGFR) allowed to identify patients who have HF.

Several acknowledged risk factors for incident HF in DM include older age, longer duration of DM, cumulative glycemic burden, higher BMI, atherosclerotic disease, elevated urinary albumin concentration, impaired renal function and hypertension [33, 34]. However, these risk factors cannot provide more personalized information about the risk of HF in a particular patient with DM because—for a clinician—it would be essential to know the exact constellation of parameters which indicate the high probability that a patient who has just been referred for the first time is at high risk of HF and should be diagnosed with this disease. Our study suggests that this answer could be obtained from the ML models. In general population of patients with DM independently of the MASLD status, the model operating on 5 features achieved high predictive performance in identifying patients with HF.

Our findings provide practical and easy to implement information about the risk factors of HF in patients with DM. Specifically, eGFR is negatively associated with the risk of HF, and others are positively associated with the HF risk, i.e., T2DM rather than T1DM, age, AF, and hyperuricemia. Several predictive models have surfaced in recent years that aimed to assess the risk of incident HF in individuals with DM, utilizing a range of factors. Williams et al. drew from electronic medical records to pinpoint predictors of HF hospitalization, including age, coronary artery disease, blood urea nitrogen, AF, and hemoglobin A1c (HbA1c), among others [33]. Meanwhile, Hippisley-Cox’s QRISK score discerned a series of parameters including systolic blood pressure, ethnicity, DM duration, type of DM and AF [34]. A post-hoc analysis of the PROactive study formulated a risk score, highlighting predictors such as age, elevated serum creatinine, diuretic use, HbA1c, duration of DM, low-density lipoprotein cholesterol, heart rate, right and left bundle branch block, microalbuminuria, previous myocardial infarction, and pioglitazone treatment [35]. Another study derived DM-CURE risk score, based on the data from the Action to Control Cardiovascular Risk in Diabetes (ACCORD) trial. It spotlighted age at T2DM diagnosis, healthcare utilization, and cardiovascular-related variables as most substantial predictors [36]. In another analysis of ACCORD trial, Segar et al. [18] developed a ML-derived (random survival forest) risk score which—similarly to us—used readily available clinical, laboratory, and additionally electrocardiographic variables. Eventually, the process of feature selection resulted in inclusion of BMI, age, hypertension, creatinine, high-density lipoprotein cholesterol, fasting plasma glucose, QRS duration, among others, as optimal predictors. Our predictive model for DM patients echoes many of these elements, including reflection of renal function by eGFR, but further introduces uric acid levels. The above mentioned models, however, were different from ours which aimed to show that the patient has the HF at the time of examination, not for predicting the future.

The subgroup analysis of patients, either with or without MASLD, revealed two distinct smaller sets of discriminative features which were subsets of 5 most important predictors elaborated for the entire cohort. This substantiates the premise that the MASLD status essentially splits patients with HF into unique phenotypes, setting a basis for specialized models that can operate on reduced number of features. These models are finely tuned for specific groups of patients delineated based on their MASLD status, and intrinsically gain from this added layer of clinical data, without the loss of classification performance. Practically speaking, this allows for a simplified approach to HF risk assessment, using fewer factors but still maintaining a similar level of accuracy. To draw a parallel, crafting specialized models for patients with known MASLD status, could be compared to the derivation of specialized models for males and females. This approach allows for selection of most predictive features that are sex-specific [37].

For the model developed specifically for patients with MASLD, only three clinical features were deemed most predictive of HF: presence of AF, hyperuricemia and reduced eGFR. All of them are associated with metabolic syndrome and both hyperuricemia and eGFR are indicative of renal dysfunction. Among the phenotype with MASLD, the cardiometabolic multimorbidity makes the presence of HF much more probable, beyond just one’s age. Interestingly, the irrelevance of type of DM likely stems from low rate of HF among those with T1DM. On the other hand, the model created only for those unaffected by MASLD demonstrates only older age and diminished eGFR as most useful to identify HF. The role of age stands pivotal, and it appears to be more telling than other clinical markers in identifying the HF. This finding is the reflection of the composite physiological alterations and their consequential impact on cardiac health, thereby serving as the main indicator for HF screening in individuals without MASLD. Supporting this observation, pooled population-based cohort study revealed that although incidence of HF substantially rises with age, and there is a significant interaction between age and established risk factors for HF such as DM, myocardial infarction, and AF. These risk factors conferred a greater risk for incident HF with either reduced or preserved ejection fraction in young compared to elderly participants. Consequently, these risk factors present a lower population attributable risk among the elderly [38].

Once again, as MASLD is a novel terminology, we have to compare our results to analysis of MAFLD and NAFLD. In the meta-analysis by Alon et al. there was a suggestion of an association of NAFLD with an increased risk of HF, AF, myocardial infarction and ischemic stroke [40] but later one by Zhou et al. highlighted a current lack of sufficient evidence to establish an association between MAFLD and AF [39]. In our study on the other hand, we indicated AF as one of the three factors enabling to detect HF among MASLD patients. Of note, eGFR was selected as a predictor in all of the abovementioned scenarios (i.e., the unknown MASLD status, and the patients with or without MASLD). These observations highlight the universality of eGFR as a marker of HF. Similarly, The Atherosclerosis Risk in Communities Study demonstrated that reduced eGFR increases the risk of incident HF among those both with and without a history of coronary heart disease [41].

Most MAFLD patients have co-existing obesity, however, MAFLD is also common in population without obesity and—among these patients—there is a higher risk of developing CVD [42]. It has been proven that non-obese MAFLD patients and patients with MAFLD and DM had a higher risk of mortality [43]. Therefore, to improve risk assessment for MASLD patients, it is important to classify subgroups within the MASLD population based on metabolic phenotypes that consider the presence of metabolic disorders. Building on this need, we focused our study on the distinctive population with both DM and MASLD. This approach could be instrumental in determining the varying levels of risk among patients with MASLD, especially when there are no approved pharmaceutical treatments and lifestyle changes remain the main MASLD treatment option that is recommended [44, 45]. Moreover, to improve both individualized management and overall public health outcomes in the context of MAFLD, two-step screening strategy combining BMI and lipid accumulation product index has been recently proposed [46].

Validation of the models proved that they discriminated the subset of patients in Dataset B with broadly similar accuracy when compared to Dataset A, showing that the models did not overfit and are able to generalize well over the unseen data (overall, the classification capabilities of the MLR models were verified over 2000 patients in total). For the scenario operating on 3 features in MASLD(+) patients and on 2 features in MASLD(−) patients, the model tended to classify more healthy patients as those having HF, while in fact they do not (hence, leading to false-positive detections). Fortunately, the predictive model for MASLD(+) individuals exhibited slightly higher accuracy in correctly classifying patients with HF. Even though there are visible differences in cohorts with and without follow-up, the model could identify patients who had HF which confirms its generalizability.

Summarizing, we automatically selected the most discriminative features from an extensive range of clinical and laboratory parameters, with the aim to enhance the precision of HF risk evaluation in patients with and without MASLD using specialized ML models. Second, we introduced a model which could identify patients with a HF risk independently of their MASLD status. This approach opens new doors toward building cascaded classification systems combining the identification of patients with MASLD and assessing their HF risk in a reproducible way. Future studies should also incorporate physical activity as one of the parameters that could modify patients prognosis since exercise increases myocardial free fatty acid oxidation in subjects with MAFLD what can be important in term of HF [47].

We are aware of several limitations of our study. This was a single center study what may limit its generalizability. We did not collect the information about the cases of gout what could be important—e.g., as additional predictors in identifying high-risk patients. We also did not analyze patients’ blood lactate levels which in a recent real-world study turned out to be associated with an increased risk of MAFLD in patients with T2DM [48]. In our diagnosis of new onset HF, we were limited by the absence of natriuretic peptide measurements, vital in cases of HF with preserved ejection fraction. Additionally, we were not able to phenotype HF based on ejection fraction, due to the lack of echocardiography data for some of the patients. Finally, although MLR models offer high-quality classification and generalized over unseen patients’ data, deploying other well-established ML techniques [49] also including deep ML classifiers [50] could further improve the classification performance of the proposed pipeline.