This study examined how task complexity affected the temporal characteristics of self-regulated learning (SRL) behaviours in clinical reasoning. Eight-eight (N = 88) medical students participated in this study. They were required to diagnose two virtual patients of varying complexity in BioWorld, an intelligent tutoring system (ITS) designed to promote medical students' clinical reasoning skills. Students' diagnostic behaviours were automatically recorded in the system log files, based on which we coded students' SRL behaviours. We used a clustering technique to classify students into high- and low-performing groups for each task. Afterward, we applied the "Fuzzy Miner" algorithm to generate SRL process maps for each performance group. We found that all students used an analytical reasoning approach to diagnose the complex patient case, whereas low performers relied on a non-analytical approach to solve the simple task. Moreover, students switched between the SRL behaviours of help-seeking and execution in the complex task but not in the simple one. Furthermore, the SRL processes of high performers were cyclically sustained in the complex task. In addition to its theoretical contribution, this study informs the design of ITS regarding providing adaptive scaffoldings and prompts in SRL.