Sepsis is a serious disease that can cause death. It is important to evaluate patientsâ sepsis risk during diagnostic decisions within the early stages after the detection of the presence of symptoms that suggest sepsis. The conventional approach to sepsis diagnosis is blood culture, which may takes several days. The approaches based on statistics and machine learning for sepsis diagnosis can be cheap, fast, and non-invasive. There are a wide variety of approaches based on statistics and machine learning that can be used for sepsis diagnosis, but these approaches have some issues, e.g. interpretability and overfitting, which may affect their performance in sepsis diagnosis. To address some of the issues in the popular approaches to disease diagnosis, we proposed a new approach, i.e., the rule-based inferential modelling and prediction. This approach integrates statistical analysis, belief rule-base inference, and maximum likelihood prediction, and machine learning. The referential-value-based data discretisation technique used in this approach is closer to reality and better at reducing information loss and distortion, as well as better at presenting the characteristics of the data, compared to other data-processing techniques. We can use the belief rule-base inference to clearly analyse the relationship between system inputs and outputs. An interdependence index is used in this approach to quantify the interdependence between input variables. An adapted genetic algorithm is used in this approach for the bilevel optimisation of models. The stopping criteria for the training process of the models used in this approach help us find the optimal structure of the models, which generally achieves balance between accuracy and complexity. Compared to the complex classifiers for disease diagnosis, e.g., ensemble, ANN, and random forest, the classifier based on the maximum likelihood evidential reasoning (MAKER) framework established by the rule-based inferential modelling and prediction approach is more interpretable. The performance of the MAKER-based classifiers constructed by this approach for sepsis diagnosis is generally better than the majority of alternative models for sepsis diagnosis, and similar to the performance of ensemble: bagged trees, which is a complex model. The MAKER-based classifier is an outstanding classifier for classical data sets: the Banana data set, Habermanâs survival data set, and the Iris data set, and it generally performs better than other interpretable classifiers, e.g., complex tree, logistic regression, and naÃ¯ve Bayes.