MIT Sloan Health Systems Initiative

Perakis: AI-driven, personalized transparent decision-support tool enables improved treatment decisions

This project addresses the question of how to find the best outcome given a situation where there are many choices in a changing environment and many possible paths. For example, how to guide a clinician to recommend the appropriate actions to treat a patient most effectively. The project lead, Prof. Georgia Perakis sums up the goal as finding the “optimal sequence of personalized treatments for patients with features that may include vitals, diagnoses, medical history and demographic or socio-economic characteristics”. To further complicate the situation:

  1. The decision maker cannot “explore the system to learn”. This is a setting where testing new actions can be costly or risky, such as in healthcare.
  2. The recommendations must be interpretable by the healthcare provider that is, the decision support tool must make a recommendation and justify that course of treatment to the clinician. This model is not a black box that spits out a recommendation.

Perakis and her colleagues Amine Bennouna (MIT), Dessislava Pachamanova (Babson), and Omar Skali Lami (MIT) are not the first to explore this problem, of course. But in a real-world application, given the number of variables and the many values they can take, constructing a usable model to account for every possible combination is often not possible or it requires very large data sets and complex calculations. Some academics have worked to construct these exhaustive models. Perakis and her team found a new approach that finds a way around these roadblocks.

What’s New

What is new about Perakis’s work is showing that constructing a usable model that can function as a decision support tool that will be used by health care providers may not require that large data set or account for every possible combination. When a clinician makes a treatment decision, not every variable such as vital statistics, or every part of a medical history, needs to be considered. Only some are pertinent to a specific decision. Furthermore, some patients are alike enough in these relevant characteristics that a clinician would make the same treatment decision for all of them at that point.  

Perakis also takes into consideration that clinicians have limited time with a patient and cannot explore freely. That is, they cannot experiment widely without considering the effect of their decisions on the patient. An additional feature of her model is that it is a true decision support tool. It provides the reasons behind recommendations, allowing clinicians more insight into the suggested treatment.

All these practical limitations are baked into Perakis’ model. Her approach makes this academic research project useful to decision makers in the real world.


Perakis tested her algorithm in a scenario of whether to prescribe short-acting insulin, considering only two patient data points: 1) last blood glucose measurement and 2) time since last insulin intake. The model’s recommendations were both justifiable and interpretable. This experiment was constrained to these two variables due to data limitations. However, the model can be further refined by working with a data set that contains information about diet, level of exercise, weight, stress level and other characteristics.  


Perakis’s research advances the fields of reinforcement learning, machine learning and statistical analysis. The practical applications of her model in healthcare are that it can recommend the optimal sequence of treatments for a patient that can be understood and trusted by the clinician. The clinician would not be working with the actual algorithm, rather this complex, robust, new method would have a user-friendly interface. The user interface shouldn’t require major new learning by clinicians but should be within their workflow and help them make better, more efficient decisions.

This model can suggest treatment paths that the clinician might not have even considered. If this happens often enough for a large number of patients, clinicians might have a new treatment path in their arsenal to better serve their patients. Over time, if the model frequently suggests specific successful treatment paths, it is possible that the standard of care could change for that condition. New “rules of thumb” might result. In both situations, healthcare providers who do not have access to the decision support tool using this model could, in this way, still benefit from it.

This article is based on the research paper:  Learning the Minimal Representation of a Dynamic System From Transition Data by Amine Bennouna (MIT), Dessislava Pachamanova (Babson), Georgia Perakis (MIT) and Omar Skali Lami (MIT)