In this study, we first demonstrated that integrated ML algorithms can be applied to predict END in AF-related stroke. Among the ML models investigated, the LightGBM had the best performance, with an AUROC value of 0.778. This is a novel method with efficient computational power and wide scalability for processing categorical, multidimensional, and incredibly large datasets,20 which makes it a suitable ML model for the medical field. In addition, this model was implemented using SHAP, which can visualize the level of contribution and directionality of specific input features using the entire dataset as well as individual patient information.
The highest contributing feature in our study was the fasting glucose level, followed by initial NIHSS score which represents the degree of initial neurological functional deficits. These variables have been consistently reported as risk factors for END in all-type as well as AF-related strokes.2,10,11 A possible explanation is that the impairment of glucose control causes vascular endothelial dysfunction,31 post-ischemic inflammatory response, and neuroprotective heat-shock chaperone gene attenuation,32 which could exacerbate post-stroke brain damage through increasing lactate production and leading to the breakdown of the blood-brain barrier, development of brain edema and hemorrhagic transformation, and enlargement of infarct volume.18 In fact, symptomatic cerebral hemorrhage of hemorrhagic transformation subtype was positively associated with END in this study. Also homocysteine, which was related to vascular endothelial dysfunction,3 and fibrin degradation product and D-dimer, which were important hematologic markers related to the coagulation system and thrombosis, were important features like previous studies.33-35 Other features were SVS presence implying large-size infarction; specific ischemic lesion location limited to anterior or posterior circulation;36 cardiac electrophysiological, and echocardiographic markers such as QRS axis, T axis and left atrium diameter; alkaline phosphatase37,38 as surrogate markers of atherosclerosis, systemic inflammation, malnutrition, or metabolic syndrome; and the burden of atherosclerosis, such as concurrent intracranial atherosclerosis.36,39 Among cholesterol lipoproteins, total cholesterol, and LDL were included as important features in this study, which have been previously reported as important predictors.40
Interestingly, the clinical implication of cut-off values in selected features may be applicable to real-world clinical practice. With regard to initial stroke severity measured using the NIHSS, cut-off values in the SHAP partial dependence plot were presented according to the effect direction of END prediction, suggesting that patients with severe stroke (NIHSS ≥ 16) tended to develop END, suggesting that awareness and close medical attention are necessary for these patients, and patients with mild to moderate stroke (NIHSS ≤ 6) have a lower chance of developing END. Some cut-off values were statistically significantly similar to clinical values. Indeed, the cut-off value for fasting glucose predicting END in our study was 117.6 mg/dL, which corresponds to the current diagnostic criteria for diabetes mellitus (≥126 mg/dL).41
The SHAP and its corresponding graphs, which were used to evaluate the effect that continuous variables had on the prediction of END, were characterized by four patterns. First, a positive correlation with or without a sigmoid or double sigmoid shape was observed. The initial glucose level, fasting glucose level, initial NIHSS score, initial mRS score, homocysteine, D-dimer, fibrin degradation product, initial diastolic blood pressure, total cholesterol, QRS-axis, and T-axis corresponded to this pattern. Most of these variables have been reported as predictors of END in previous studies.2 Second, a U-curve or J-shaped pattern with both cut-off values was observed for aspartate aminotransferase, alkaline phosphatase, total bilirubin, and LDL cholesterol. Under the lower cut-off value of each feature may have been associated with poor nutritional status and over the upper cut-off value may imply comorbid conditions including liver disease and hyperlipidemia. However, it is unavailable to investigate underlying pathomechanism of these phenomena in this study. Third, the following had a negative correlation with END, with a reverse S or J shape: LA diameter and uric acid. In particular, the negative association between LA diameter and END is not consistent with the positive correlation found in a previous report.42 However, more accurate parameters, such as the LA volume index, have recently been identified as important predictors. Considerable imputation (21.1%) could lead to incorrect directions and biased results. Finally, a bizarre pattern with multi-directionality was observed in activated partial thromboplastin time.
One strength of our study is that our interpretable ML model was constructed using many variables, including demographics and laboratory, radiological, and echocardiographic findings, all of which can be obtained upon arrival at the hospital. Additionally, an interpretable and explainable ML model was created to promote the use of applications for making clinical decisions. Our study demonstrates the potential of interpretable ML methods to predict END and individualize such predictions. Previous studies have focused on each risk factor individually and its pathophysiological interpretation, but there has been a shortage of clinical use of a large combination of variables at once.3-5 Moreover, no standardized risk stratification scheme for predicting END has been available until now. Therefore, our ML model has the advantage of being able to predict END using diverse variables extracted from real-world clinical situations upon arrival at the hospital.
Our study has several limitations. First, a considerable amount of data was missing due to the multi-center retrospective nature of the study. Although imputation of missing data was performed using the ML technique, the results may be biased and contradict previous findings. In particular, it seemed to occur with some elements (such as left atrial size) that were less important. In addition, laboratory and imaging protocols in each center were not concretely established before data collection. Additionally, Holter and electrocardiography parameters were not standardized; therefore, many variables were excluded. Second, since this was a registry-based study with a retrospective design, the ML model’s performance is not sufficient to be an absolute criterion for clinical use. It is necessary to conduct prospective studies, develop a more accurate prediction model, and discover novel biomarkers for a deeper understanding of the pathophysiology, in parallel. Third, the implementation and evaluation of the model was difficult to generalize because of the lack of external validation. Further verification is required through well-designed prospective clinical studies in the future.
In conclusion, ML algorithms, using the LightGBM model in particular, can be used to predict END in AF-related stroke. New and diverse predictors for END were revealed through this ML model, suggesting that the pathophysiology of END development could be a complex mechanism. Further verification through prospective clinical studies is needed.