AI and Reinforcement Learning in Precision Medicine

Author(s) : Shyamsree Nandi
ISSN : 0974 - 497
Year : February 2019 | Volume : 13 | Issue : 1/4

Abstract: Healthcare and Life Sciences have begun realising disruptive innovations using a combination of Big Data, AI and Machine Learning. One of the big motivations of this industry for the past couple of decades has been the need to personalise medicine and treatment. Whether we talk about "value-based care" or novel treatment approaches specific to an individual's genome pattern, the use case that is subtly underlined across all of these subjects is about measuring, predicting and self-learning models for treatment pathways and associated outcomes. The estimated market size of personalised medicine worldwide is projected at 2.77 trillion U.S. dollars by the year 2022 at a steady CAGR of about 12%.

Precision or Stratified Medicine holds great promise for humankind and has seen tremendous adoption rates due to the fact that it leads to tailored interventions and hence much better outcomes across different patient groups. That said, there is a significant amount of challenge in scientifically arriving at personalised treatment plans. Can Big Data and AI be the fundamental backbone of such methods? In this paper, we will explore the possibilities of using a sophisticated Reinforcement Learning technique viz. Markov Decision Process (MDP) on datasets residing in a Data Lake to guide stratified medicine. This solution helps in achieving a better response, higher safety margins and lower treatment costs.

MDP is uniquely suited to the parlance of medical science. At each stage of a disease progression the doctor has a set of actions to choose from and based on the action chosen, the patient is transitioned to another state. A sequence of such states may eventually lead to either a positive or negative outcome. At any given point in time, the desired outcome of the treatment/ intervention will be to maximise the reward or chances of achieving a positive outcome. Patient Data collected across different states of the disease can be consolidated, curated and standardised in a Big Data Lake. MDP can then be used across hundreds of genomic sequences to identify the actions that can lead to the best treatment outcome.

The fundamental gap that exists in care delivery today is the lack of a guided mechanism to assist doctors' decision of choosing appropriate treatment pathways. The methodology involves analysis of clinical datasets that provide necessary information to trace longitudinal view of diseases, followed by an architecture view of how to store and organise datasets, and disease-specific MDP modelling. Leveraging AI the models adapt over time leading to a more pragmatic guess about an intervention that is likely to result in the best possible outcome for a patient with specific clinical features and disease history.

Connecting all past events in a patient's medical journey with future outcomes is the uniqueness of MDP driven clinical decision support system and hence when coupled with the experience of a practitioner leads to a holistic analysis of expected outcomes ahead of time. Reinforcement Learning within AI thus offers an alleyway to maximise care delivery performance with far-reaching implications for the society.


Download File Download File Open File