According to the National Center for Chronic Disease Prevention and Health Promotion, six in 10 adults in the U.S. have a chronic disease, and four in 10 adults have two or more. Combined with mental health care, treatment for chronic diseases accounts for 90% of the country’s $3.3 trillion in annual health care costs. It’s no wonder that so many initiatives are in flight that are taking aim at prevention and management of chronic disease.

Given the complexity of these endeavors, it’s also little wonder that stakeholders are turning to artificial intelligence (AI) and algorithms to help. Consider that the trajectory of chronic disease is different for every individual and is dependent on many factors such as age, gender, ethnicity, race, genetics, lifestyle, dietary habits and pre-existing medical conditions. The rise of precision medicine and the increased technological feasibility of gathering and storing genomic, transcriptomic and proteinomic data (known as “–omics data”) has made it a necessity to turn to digital phenotyping – the quantification of individual-level behavior using data from personal digital devices – to understand disease progression at an individual level. Digital phenotyping, in turn, has spurred the use of collaborative data networks that facilitate sharing of electronic medical record (EMR) data among participating entities.

All of these factors combined point to the need to use algorithms to understand disease progression at an individual level. Needless to say, the major industry stewards indeed are working toward this goal.

Meanwhile, physicians face the mammoth challenge to decipher the right treatment considering an ever increasing body of medical research, lack of historical precedents for all treatment pathways across all different types of phenotypes, cultural differences in various population segments, inherent socioeconomic biases in available medical research data, increased globalization and delayed data on the effects of treatment administered.

Clearly, there’s an eminent need for a clinical decision support system that specifically focuses on slowing chronic disease progression and prescribes the targeted and specific next best action for effective management of chronic disease, at an individual level.

Introduction to Evolutionary AI

An advanced form of AI, Cognizant’s Evolutionary AI^TM, is well-suited for this mammoth task, as it allows thousands of possible solutions to be evaluated in parallel and selects the best one. It is a self-evolving system that learns at scale using machine discovery and principles of biological evolution. At its core, Evolutionary AI uses principles of genetic programming and biological evolution, including reproduction, mutation, re-combination and selection. It picks solutions that have worked and, using cross-over and mutation, generates new solutions, that would not have been feasible through traditional machine learning methods because of lack of historical data on success/ failure for all permutations and combinations, amongst other reasons.

A key point to understand is that Evolutionary AI facilitates machine discovery at scale, and establishes a system for rapid proto-typing in a business-as-usual setting. After the initial setup this system doesn’t need significant human intervention, and can evolve by itself at a rapid pace, using principles of biological evolution as new data, new feedback and real world evidence flows in. This is key imperative for any AI system in clinical decision support capacity to succeed. After all, now more than ever, technology and advances in medical science are making it feasible to collect, store and analyze ever expanding -omics data.

Importantly, this system can be bootstrapped with any existing algorithm. This is critical because it enables us to leverage previously developed algorithms (by industry and academia) to understand chronic disease progression. A lot of work has been and continues to be done through joint collaboration of academia and clinical setups to further the development of these algorithms, and it just makes more sense to start with something already developed and build on that.

Building an Evolutionary AI system involves two key steps:

• Building a predictor surrogate (simulation) model. A predictor model is a standard machine learning model that takes context data (specific to the problem) and historical actions data as input, and attempts to predict the business outcome.
• Training a prescriptor model. A prescriptor is a model developed using deep learning techniques, and is based on principles of biological evolution using evolutionary computation. This model will generate prescriptions on what action should be taken, to optimize certain business outcomes. With Cognizant’s Evolutionary AI system, this is done through an API call. This can’t be done using standard ML techniques because the optimal solution isn’t always known; it requires machine discovery.

Slowing Chronic Disease Progression

Let’s look at the case of chronic kidney disease (CKD) and how evolutionary computation can slow its progression (see diagram below).

Let’s talk about all the blocks now and what they do.

• Predictor: The predictor in this case is the algorithm that will understand the progression of chronic disease using digital phenotypes. This algorithm shouldn’t be built from scratch since algorithms already exist for extracting digital phenotypes using EMR data from collaboration networks like DARTNet and eMERGE. Several other existing algorithms can also be used to bootstrap the predictor. It is advisable, however, to enhance the chosen model to include attributes and data for interventions, such as treatment decisions made, for the population in question.
• Prescriptor: Using principles of evolutionary computing and the power of machine discovery, the predictor is used as an input from which a prescriptor is developed to recommend possible interventions to slow the progression of CKD. The system itself is not taking action; it serves as an input to the physicians to support and enhance their clinical decision making.
• Physician: Physicians can now use their clinical experience and expertise, the specific knowledge gained through interaction with the patient and guidance from the prescriptor to determine the best course of action for the patient. The physician also documents actions taken and not taken and the corresponding results and reasons. With time, as the evolutionary computation system learns and becomes more intelligent, it can serve as an intelligent assistant to support the physician’s decision-making process.
• Patient: The role of the patient is critical to the evolution of the system because this evolution is dependent on the patient adhering to physician recommendations and reporting the results back accurately. Clinical tests will certainly help validate the clinical efficacy of the intervention, but the accuracy of the patient in reporting his or her degree of adherence to recommendations is critical.
• Feedback System: This system has two purposes: capture the results of recommended interventions in terms of clinical efficacy, and search for new training data for the algorithm, such as EMR data. This system would then pass on this information to the predictor, from which the predictor and prescriptor would learn. For the latter, the system would continuously monitor the collaborative data network inventory (like DART and eMERGE) for new available training data. It can be enhanced to monitor other data sources as well.

One key point to remember is that in healthcare, the cost of trial and error could be a human life; for evolutionary computation or any other AI system to succeed, collaboration between humans and machines is hence a must.

The tasks outlined above are far from simple, and the solution won’t happen overnight. However, Evolutionary AI does show a promising way forward to understanding disease progression at an individual level, and introduces a real opportunity to slow down disease progression through targeted intervention design.

To learn more about Cognizant Evolutionary AI, visit us at our website.

 

Kapila Monga is an Associate Director in Cognizant’s Digital Business AI & Analytics Practice and focuses on designing and delivering machine learning and AI solutions for healthcare payers and providers. She has led teams to deliver AI, ML and advanced data science solutions that have yielded over 150 times the return on investment.

Kapila is a regular contributor in the Data Revolution blog and the Illuminating Informatics column of the Journal of AHIMA on topics related to the application of machine learning in healthcare. She is a firm believer in the promise of artificial intelligence for healthcare and intends to work with healthcare clients to see this translate to reality.

Kapila has over 12 years of experience and holds a master’s degree in mathematics from the Indian Institute of Technology (IIT) Delhi as well as an MBA in strategy and leadership from the Indian School of Business, Hyderabad, India. She can be reached at Kapila.monga@cognizant.com