Fig. 6: Predicted disease course at 93% confidence for an example patient with 13 hospital visits complemented with XAI.

a Disease course of a transitioning patient with RRMS at the initial hospital visit and SPMS at the final hospital visit. In disease course plot 1 (top figure), the model predicts that the transition occurred in year 3.8. However, the clinic’s SPMS assessment could not be made until between year 3.8 and year 7.1. Thus, as indicated by the gray zone, a delay of three years is observed. Thus, the model identifies SPMS early, approximately three years in advance. Disease-modifying treatment names taken during the disease course are listed atop the figure. The disease course plot 2 (bottom figure) manifests the progression of the disease towards SPMS, indicating the disease worsening over time. A clear drop in RRMS p-value occurs between years 0.7 and 3.8, and at the same time, an increasing p-value score for SPMS is observed (between years 0.7 and 7.1). As the disability accumulates, the plot illustrates a decreasing RRMS p-value with an increasing SPMS p-value. b Feature contribution explanation using force plots for the predictions on the hospital visit at years 0.7, 3.8, and 6.5 of the patient. During the hospital visit year 0.7, the model predicted RRMS, driven by lower EDSS score, first-line DMT, and age at the visit. Meanwhile, the features contributing to SPMS are minimal. Conversely, in year 3.8, the model predicted SPMS, influenced by factors such as age at relapse, number of steroid treatments, age at SDMT, and lack of steroid treatment. Features such as age at visit, first-line DMT, EDSS score, and age at MRI contributed towards RRMS. By year 6.5, a high EDSS score considerably influenced the prediction of SPMS, while first-line DMT and age at visit contributed to RRMS. (The results of all visits are found in Supplementary Figs. 9–11).