Artificial intelligence (AI) is popular for complex decision-making analysis, but interpretation is challenging. Interpretable AI is a flourishing research field that builds ML complete solutions by bridging the gap between performance and interpretability [24, 41]. SVM is a popular ML model known for producing accurate predictions in classification problems [42]. It generates a multidimensional hyperplane that distinguishes classes [33], maximizing the margin between observations and minimizing training errors [34]. SVM has high discriminative ability in small sample sizes and large variables [43,44,45], and good generalization ability, robustness, and avoidance of overfitting [28], but has difficulty in interpreting results [29]. To the best of our knowledge, two studies have been identified as interpretable AI fields based on the SVM model for healthcare research, i.e., [20] used a hybrid approach based on Quadtree decomposition to interpret the results of medical image analysis, and [21] used local ICE interpretability techniques to interpret the results of marketing research that classified customer preferences. Moreover, [30] utilized global and local interpretability techniques to interpret hypertension prediction outcomes using the RF model. Global interpretability techniques generalize over the entire population, while local interpretability techniques provide instances-level explanations. Both methods are computationally expensive, but valid depending on application needs, such as healthcare applications [30]. Therefore, this research aimed to predict infant mortality in Bangladesh based on the BDHS 2017/18 data using global surrogate models and local ICE interpretability techniques, enhancing clinicians’ understanding and trust in healthcare analytics at local and global levels.

Age at first marriage, Successive Birth interval, Division of residence, Mother’s educational level, Father’s educational level, Wealth index, Mother’s BMI, Religion, Respondent occupation, Exposure to mass media, Gender of child, Birth order number, Toilet facility, Type of cooking fuel, and Total children ever born are the variables that the chi-square test identified as significantly associated with infant mortality in Table 1. In contrast, the most important characteristics to predict infant mortality in Bangladesh are identified by SVM-based feature selection as being the type of cooking fuel, successive birth interval, mother’s occupation, place of residence, religion, toilet facility, birth order number, TT vaccination, mother’s educational level, child’s gender, and mother’s BMI. We estimated both the LR model and SVM model with the hyperparameter tuning parameters for the entire dataset to predict infant mortality in Bangladesh using these selected predictors (Fig. 2). Based on the prediction errors, the LR model had an AIC value of −65,110.68, while the SVM model with the sigmoid kernel had an AIC value of −50,559.46 (Table 3). Additionally, the LR model provided a superior fit based on the RMSE and BIC values. The study analyzed the effectiveness of machine learning classifiers in predicting infant mortality in Bangladesh. The dataset was divided into 70% training data and 30% test data, with 100 permutations conducted. The results of the 100 permutations demonstrated that the LR model (Average: accuracy = 0.9105, precision = NaN, sensitivity = 0, specificity = 1, F1-score = 0, area under the ROC curve (AUC) = 0.6780, run-time = 0.0832) outperformed the SVM model (Average: accuracy = 0.8470, precision = 0.1062, sensitivity = 0.0949, specificity = 0.9209, F1-score = 0.1000, AUC = 0.5632, run-time = 0.0254) in predicting infant mortality (Table 5), but the LR model had a slower run-time and it was unable to predict any positive cases. The test data set for random seed 1980 had 671 positive cases and 7173 negative cases, but the LR model was predicted the true negatives (TN) cases were 7173, the false negatives (FN) cases were 671, the false positives (FP) cases were 0, and the true positives (TP) cases were 0 in Table 4. Hence, the LR was unable to predict any positive cases. However, the SVM model with the sigmoid kernel was able to identify some positive cases, these include TN (6630), FN (611), FP (543), and TP (60) in Table 4.

The LR analysis found that infants were less likely to die when mothers gave birth after a gap of over two years (Table 6). Mothers with higher educational attainment had lower infant death rates. Normal BMI mothers had a higher risk of infant death compared to underweight mothers. Overweight or obese mothers had a lower risk. Working mothers had lower infant mortality rates. Male infants had higher death rates than female infants. Infants in positions two or higher had lower mortality rates. Families with polluted cooking fuel had less infant mortality (Table 6).

The study examines Bangladesh’s infant mortality prediction based on the SVM model using the local ICE interpretability technique (Fig. 6), which depicts individual influences on the average likelihood of dying within the first birthday. The study results reveal that mothers with normal BMIs, less polluted cooking fuel, and male newborns are more likely to experience infant death. Working mothers are more likely to experience infant deaths than non-working mothers. Infant deaths are more common in non-Muslim families and those who give birth within two years of previous birth. Hygienic toilets and vaccinations are also less likely to cause infant deaths. The northern administrative divisions of Bangladesh have higher infant death rates (Fig. 6). The global surrogate model specifies the correlation between a particular predictor (with the remaining predictors left unchanged) and the likelihood of seeing an infant’s death. A fitted SVM model using the global surrogate model predicts 2296 infant deaths and 23,849 alive, with a higher number of infant deaths occurring when households use polluted cooking fuel in Fig. 7. The model also predicts more infant deaths for working mothers who also use polluted cooking fuel in their households, given that they use less polluted fuel and have a birth interval of more than two years. Mothers who give birth within two years of the previous birth also have more infant deaths predicted by the fitted SVM (Fig. 7).

Our findings demonstrate that the type of cooking fuel has a higher significant impact on infant mortality, even when considered alongside other variables such as Successive birth interval and Mother occupation. For instance, mothers who maintained successive birth intervals of less than two years experienced higher child mortality rates when exposed to polluted cooking fuels, which shows that cooking fuel consistently affects them. Evidence from several studies supported our findings that polluted or solid fuels caused more infant deaths [15, 16, 46, 47]. The global surrogate model reveals that mothers who lived in less polluted areas and used clean fuels and hygienic restrooms had lower death rates regardless of other factors. Our study provided strong evidence that a lower birth interval (less than two years) is more responsible for increasing the risk of infant deaths, which is supported by the earlier study [15, 16, 48]. This study also found working mothers, parental poor educational qualifications, reluctance to take the TT vaccine, gender of child and unhygienic toilet facilities have significant impacts on increasing infant deaths. The LR model indicated that working mothers and houses with contaminated cooking fuel had lower infant death rates (Table 6), which fails to reveal the original situation. However, the global surrogate model predicts a higher infant mortality rate among working mothers who use polluted cooking fuel at home (Fig. 7).

Complex ML models, while outperforming simple interpretable models, clinicians struggle to understand and trust these complex ML models due to their lack of intuition and explanation of their predictions. By utilizing the global surrogate model and the local ICE interpretability technique, we were able to predict infant mortality in Bangladesh and interpret the findings of the SVM models based on the BDHS 2017–18 data. This improved clinicians’ understanding and trust in healthcare analytics, enabling them to take further initiatives and interventions for stakeholders and policymakers. One of our study’s limitations is that we could have improved our results by using the globa surrogate model with terminal nodes of depth equal to 2–8 instead of 2–4 and some more other interpretability techniques. Furthermore, the cross-sectional BDHS data, which is derived from a national survey, might contain some biases. These could be addressed as a study limitation. However, our developed interpretable SVM model reveals global interpretations help clinicians understand the entire conditional distribution, while local interpretations focus on specific instances, providing different insights into model behavior.