In collaboration with a number of units of the Royal London Hospital, Barts and The London School of Medicine and Dentistry, as well as other medical practitioners, the Risk Information Management research group (RIM) has developed a number of intelligent causal models for improved medical decision making. These models – based on Bayesian Networks (BNs) include: predicting risk of strokes; coagulopathy risk; limb viability after traumatic event; musclo-skeletal triage and sports injuries, managing recidivism in violent psychiatric patients; managing Warfarin therapy). The underlying models combine four different types of data:
- Historical data relevant to the general medical condition(s) and its treatment
- Expert medical judgment – including causal relationships between relevant factors – about the general medical condition(s) and its treatment (which is paramount, as in many situations the relevant historical data is extremely limited, and many key factors may be unobservable).
- Background patient-specific data
- “Live” data from patient examination and clinical tests
The first two types of data are the basis for the pre-defined model (which is called ‘expert driven’ because of the inclusion of data type 2), while the last two are the core model inputs for a given patient. The models are probabilistic: when the patient specific inputs are entered into the model, Bayesian inference computes updated probabilities for all ‘unobserved’ variables of interest (such as the probability that the patient: will suffer a particular medical event within the next x hours; has a particular condition; will respond to a particular treatment).
The resulting ‘decision support systems’ have proven accurate and effective when compared to previous state-of-the-art methods that are purely data driven (this includes data-driven Bayesian methods that work well in epidemiological applications but do not work in the kind of applications described above where there is limited relevant data).