We used an interpretable ML statistical framework and the Kuwait CLAP registry data to reveal deeper insights into the risk factors that shape the outcomes of admitted patients with ACS during hospital stay and at 30 days from discharge. Also, we demonstrated how our ML analytical pipeline could untangle the unique and complex relationships between the different risk factors related to patient characteristics and in-hospital clinical procedures. Also, we showed that our most important predicting features had remarkable non-linear relationships with other baseline characteristics in shaping the risk of clinical outcomes. These findings not only support and improve clinical practice but assist with alleviating the public health and economic implications of ACSs.

Our in-hospital adverse events ML model identified LVEF as the most important risk factor (Fig. 1) with the highest interaction strength with other features (Fig. 3). This is unsurprising since many past studies highlighted the critical role of low LVEF values in influencing the risk of post-catheterization in-hospital adverse events^36,37,38,39. Here, our cICE plot demonstrated that LVEF values on the admission of less than 40% increase the risk of post-catheterization in-hospital events (Fig. 2A). However, this risk is aggravated particularly in older patients (Figs. 2F and 3E) with heart failure (Figs. 2C and 3F), high RVSP (Fig. 2D) and irregular systolic blood pressure (Fig. 2E). All these features represent severe cardiac insufficiency leading to poor long-term prognosis, and thus, need to be taken into consideration when performing any catheterization procedure in ACS patients³⁹. Further, our results showed that patients receiving furosemide within 24 h of admission are at elevated risk of in-hospital adverse events (Fig. 2B).

Also, interaction plots illustrated a significant non-linear relationship between low LVEF values and receiving furosemide in amplifying the risk of adverse events (Fig. 3C)⁴⁰. Furosemide is commonly used as a loop diuretic for patients with heart failure. Therefore, our results reflect that patients with severe cardiac outcomes (such as low LVEF) requiring a high dose of furosemide may have a poor prognosis⁴¹. The same may also be implied in patients receiving high doses of aldosterone within 24 h of their admission, as shown in Fig. 3D. Moreover, adverse events may also result from the rare side effect of these medications, thus combining them with other medications may improve their therapeutic outcomes⁴². Additionally, our model not only demonstrated that patients with chronic renal failure and diabetes are at high risk of adverse events as suggested elsewhere^38,43,44,45, but revealed significant interactions with low LVEF values (Fig. 3G,H). These findings agree with the notion that the combination of hyperglycemia and renal insufficiency associated with low LVEF values is the leading cause of in-hospital adverse events, particularly in patients who have undergone a PCI operation³⁸. Also, these poor outcomes might reflect the low cardiac output, hemodynamic instability, and reduced renal blood flow, which leads to hypoxia and the generation of reactive oxygen species⁴⁶.

Nevertheless, the 30-days adverse event model inferred that highly invasive intervention procedures, such as performing urgent CABG, having multiple culprit arteries, with stents placed during PCI, are significant predictors of poor outcomes after discharge (Fig. 2G–I). These findings confirm the results of past studies in terms of reflecting the severity of the patient’s ACS condition^16,47. Also, this is evidenced by the importance of the cardiological and hematological indicators such as RVSP and platelets, respectively (Fig. 2J,K), as well as having an in-hospital adverse event (Fig. 2L). However, unlike previous inferences^3,6,47, our model uncovered the strong non-linear relationships between admission creatinine levels (Fig. 4A) and other features (Fig. 4B) in shaping the risk of adverse events after discharge (Fig. 4A). Here, our inferences demonstrate that patients requiring urgent CABG with creatinine levels less than 50 µmol/L or greater than 100 µmol/L are more likely to experience a poor post-operative prognosis (Fig. 3C). This result is expected since abnormal serum creatinine levels correspond to other comorbidities, particularly chronic kidney disease, exacerbating the long-term risk of postoperative adverse events⁴⁸. Similarly, serum creatinine had a strong non-linear relationship with ACEI and ARB intake after discharge (Fig. 4D,E) in hypertensive patients.

Nonetheless, our results show minor discrimination in the risk between patients discharged with and without ACEI (Fig. 4D). In contrast, remarkable discrimination was inferred between patients discharged with and without ARB medication (Fig. 4E). These findings quantify the notion that ARB may increase the risk of myocardial infarction (MI) in hypertensive patients, and therefore, dispensing ACEI to control their blood pressure may be more appropriate, particularly for acute MI patients, as suggested elsewhere⁴⁹. Also, our model was able to discriminate the broad spectrum of risk of poor outcomes among diabetic patients with abnormal serum creatinine levels (Fig. 4F). These results suggest that severely diabetic patients (i.e., who are under insulin injection as a proxy) are more likely to experience adverse events than moderately diabetic patients (i.e., who are under oral medication). Indeed, the complex angiographic pattern extending between the mid and distal arteries of ACS patients with severe diabetes is characterized by a multivessel diffuse plaque, making revascularization quite challenging for clinicians⁵⁰. Thus, interventional cardiologists and cardiothoracic surgeons might need to implement an individualized approach with a multidisciplinary heart team on severely diabetic patients to minimize poor outcomes after discharge⁵⁰.

One limitation of this study is the aggregation of positive outcomes into one category in our cohort. Yet, the rarity, complexity and broad spectrum of outcomes (Table 1) made it difficult for us to generate a representative model for each adverse event. However, the aggregation of the adverse events increased our computational efficiency, substantially improved the predictive performance of our ML algorithms, and facilitated the practical interpretation of our models. A second limitation of the Kuwait CLAP registry is the population size, and therefore generalizability of our inferences might be biased toward the population that comprised our analyses. That said, many of our findings agree with past studies regarding short- and long-term adverse events resulting from post-catheterization. Furthermore, our analysis mainly focuses on revealing complex relationships in the available data that might be useful for improving clinical decision-making related to the diagnostic and prognostic efforts in the same population where the data were retrieved. This is in addition to the fact that data is being collected from only sites that provide cardiological services in the country, as described above, making it representative of the whole population of Kuwait. Also, our k-fold cross-validation procedure lessens the chances of overfitting, increasing the robustness of its subsequent inferences. Nevertheless, future studies will be aimed at applying our analytical pipeline on a larger sample size and will be focused on building specific models for the most prevalent adverse events.

The complexity of ACSs epidemiology, the growing volume of cardiac intervention procedures with their related data, and the highly non-linear relationships between patient baseline characteristics, clinical procedures, and interventions highlight the utility and robustness of our ML statistical framework. One important highlight of our analytical pipeline is the ability to flexibly explore heterogeneous treatment effects (i.e., effect modification and beyond) comprising multiple features simultaneously rather than overall average intervention effects using one-way or more interaction terms as in traditional regression models⁵¹. Due to the tedious task of modelling and interpreting all possible interaction terms, rigorous evaluation of heterogeneous treatment effects has yet to be widely explored in clinical epidemiology⁵². As shown in Figs. 3 and 4, investigators can intuitively interrogate multiple interactions to capture clusters of subgroups showing different feature-outcome effects. For example, Fig. 3G simultaneously shows how the risk of adverse in-hospital events has distinct patterns of over six significant interactions. In these interactions, the highest risk of adverse events notably peaks over certain clusters of patients with specific interrelated features (Fig. 3C–H). This allows clinicians to assess the effectiveness of their interventions and formulation of targeted approaches for reducing cardiovascular adverse events for individual clusters of patients. Wiemken and Kelley., 2020 extensively discussed the advantage of the ML algorithmic approach in dealing with interactions, as well as how traditional stratified regression models and the inclusion of interaction terms can lead epidemiologists to the issue of multiple testing bias⁵¹.

Here, our ML models had good and similar predictive performance compared to past studies in terms of evaluation parameters (e.g., AUCs = 0.84 & 0.79 for the in-hospital and 30-day adverse events models, respectively, Table 2)^{7,14,53,54,55}. Further, we showed how RF and XGM algorithms can remarkably outperform traditional models such as logistic regression (Table 2). Subsequently, many studies also demonstrated that our statistical approach outperforms standard risk stratification tools such as TIMI and GRACE^7,17,54. However, many of these ML studies mainly focused on their models’ predictive power (i.e., using a black-box approach), which they did not embrace their interpretability in a clinical setting. Hence, a readily interpretable model will provide new insights into the complex epidemiology of adverse events and be easily adopted by cardiologists to be implemented in their practice. Given that the Middle East has the highest incidence of CADs on a global scale⁴, our study represents the first attempt to utilize an interpretable ML statistical framework focused on uncovering complex relationships to improve clinicians’ intervention efforts.

Besides the inherent limitations of the statistical framework used to build standard risk stratification tools, the generalizability of their inferences might also be restricted to specific populations. Indeed, the environmental, genetic, and clinical settings and resources might differ substantially between countries and regions worldwide. Thus, a customized risk stratification tool based on local data will provide more plausible and generalizable inferences for its source population than global-based tools. Therefore, we further elucidate the remarkable applicability of Shapley values, a game theoretic approximation, to interrogate in finer scales what each model represents regarding the predicted risk of adverse events (e.g., why a particular patient had a poor post-catheterization outcome, while the other did not?). For example, the in-hospital model inferred remarkably different magnitudes of risk for different types of adverse events in individual patients instead of averaging over the risk profiles of these patients (Fig. 5). Here, our model predicted high probabilities for specific adverse events (P = 0.79; Fig. 5A), such as in-hospital heart failure and contrast-induced nephropathy in older patients with chronic renal failure who had an urgent CABG. However, midrange probabilities were predicted for other adverse events, such as acute thrombosis (P = 0.28; Fig. 5B), in younger patients with prior CVA and who had a basic PCI. Thus, both types of patients had notably distinct demographics and clinical features with different requirements for in-hospital procedures. Additionally, for a randomly selected patient who had an adverse event 30 days after discharge, having an urgent CABG with multiple culprit arteries and stents placed during PCI put that patient at high risk of having a poor outcome (i.e., 72% chance; Fig. 6A). In contrast, under the same predictive model, the other selected patient who had no adverse events 30 days after discharge, entirely lacks such risk profile (Fig. 6B).

Finally, the Shapely statistical procedure assigns positive and negative values for the features that increased and/or decreased the probability of adverse events in individual patients, respectively (Figs. 5 and 6). Hence, using such an intuitive approach can provide additional guidance to the clinician’s diagnostic and prognostic efforts and aid in allocating intervention resources to patients at higher risk, whether in-hospital or after discharge. Yet, additional evaluation of the technical feasibility and clinical plausibility are crucial steps before integrating such predictive models into the standard healthcare systems¹⁵.