top of page

Dial Dr Watson? AI & Machine Learning in Healthcare

Alan woke up after a troubled night. Panda his virtual medical assistant had already sensed that. Panda noticed Alan had a restless sleep and his vitals were far from normal. His galvanic skin response indicated stress, blood pressures were abnormally high and to top it he had mild fever. Panda in the meantime had scanned Alan's genetic, proteomic and his clinical data- and past history. Panda matched the symptoms with a cohort of like patients and was ready with a prediction. Alan needed a medical intervention. Panda had "virtually alerted" the hospital and an ambulance was on its way. In the meantime Panda had created a case summary of Alan's health, his clinical journey, interventions in the past and the outcomes, transmitted Alan's vitals for the past 30 minutes. Panda had also provided a prediction of likely episode based on data that it had analyzed in a semantic data base and a recommendation on a treatment protocol personalized to Alan. Panda had also alerted Alan's daughter.

Fiction? Welcome to the future world of AI and machine learning.

AI and Machine Learning stirs strong emotions, some are extremely skeptical some fear it and some are plain excited. Healthcare is a complex discipline. People with similar disease respond differently to similar treatment plans. Physicians are "investigative clinical analysts" who dig deep into clinical data, derive structure from the data (patient treatment response) and adapt treatment plans as new data comes in. This is precisely how machine learning works and Healthcare is one vertical that is embracing AI. Accenture estimates that the Healthcare vertical can see an annual savings of over $150 Billion a year through the usage of AI and machine learning technologies.

In fact the AI health market is witness to a dramatic growth. Most technology companies - IBM, Google, Amazon, HP are doubling down on their investments in Artificial Intelligence, Analytics and Machine Learning for Healthcare applications. Accenture estimates that the AI market will register a compound annual growth rate of over 40% through 2021.

So, why is Healthcare ripe for adopting AI and Machine Learning?

Let's look at the physical process of diagnosis.

Alan set's up an appointment with his physician. His commute could be anywhere between 30 minutes to an hour to the clinic, his wait time to see his physician even longer. He gets his vitals taken by a nurse who records this (fortunately) directly into the EMR. His physician arrives and inquires from Alan the reason for the visit. Alan's past history is recorded in the EMR, 30-40 pages, but his physician does not have the time to review that entire history, after all each page would take a minute to read. He examines Alan and has an inkling on what could be wrong based on symptoms, initial investigation and past experience. But this needs to be confirmed and he needs more data. Alan may go in for a stress test, EKG, MRI or even some pathology. The doctor now has a lot more data- he learns structure from data- and creates a preliminary diagnosis.

Based on past learned experience and expertise (his neural net) he decides on an initial treatment plan. In his "judgment" this would be the most "probable" pathway to success. Alan's doctor has limited information available to process and his prediction or judgment is no different to a rules based expert system. Alan would be provided a treatment regiment, placed under observation and his responses to the clinical actions measured, while the care team and his physician solve instances that are natural and have structure (Discharge planning process). The intermediate outcomes feed back to the doctor who refines his predictions and his treatment plan. (A trial and error framework, improving accuracy and the direction of improvement)

Not everybody responds in the same way. Alan's condition may be complex and his physician may call a clinical conference of specialists to discuss all the "Data" from Alan's situation. "Swarm intelligence" takes over and the neural networks enlarged and become stronger (Neural nets for deeper problems). Collective wisdom provides a refined judgment and a revised treatment plan (Learning structure from data and adapting as new data comes in) . The process repeats, at each step there is more data to work on, the probability of diagnosis improves and Alan would soon be ready to go home to recover.

How would AI and Machine Learning change this scenario?

Healthcare is a big data problem. The human body is complex- Proteomic, biomic, genetic makeup with over 3000 metabolic pathways to delve into, add to that the clinical data and then the high velocity daily physiological data populates just a very small portion of the semantic data lake so useful in diagnosis. Medical research is producing tons of papers body knowledge in conventional therapies, new therapies immunology and DNA synthesis.

No physician can process so much data- which is why healthcare still operates on a rules based expert system framework.

Machine learning requires large data sets. It is algorithmic based and solves for instance that are natural and have structure. Machine learning relies on a trial and error framework, trained on a training set of data and as more data and "neural nets" deepen the training accuracy improves.

Take a set of MRI images some with a tumor some without. The machine learning algorithm scans the images and is trained to identify tumors. If there is an error in identification, that is fed back to the algorithm which then refines its algorithms to better predict. The larger the volume of images to be analyzed and the diversity of observed tumors, the better the algorithm predicts a tumor. Over a period of time it can detect with near certainty and sometimes tumors that escape a human eye. Similarly a cognitive computer could be trained to study angiograms to detect early signs of congestive heart failure.

IBM's Watson is in the forefront and has built collaborative partnerships with leading health systems to use AI and machine learning to provide better assistance to physicians.

In an AI model, Alan's clinical, biomic , physiological data would be in a large semantic data lake. There would be logical connections to his siblings and parents data as well as to a cohort of patients with similar condition (Deep neural networks). Symptoms would be input to the AI system and based on history, existing medical research, body knowledge the algorithm could predict the most probable condition- with adequate backup information to justify as to how it reached that conclusion and then guide a physician to make a call on the most appropriate treatment direction. In fact over a period of time and with enough data the model could recommend a treatment plan.

In this revised scenario Alan would transmit his symptoms, along with his vitals to a Cognitive platform in the cloud- almost any vitals captured in a clinic can now be captured on Smartphones- portable ultrasounds on smartphones may not even necessitate steth readings.

The cognitive computer would analyze current symptoms with past history and all available data from his EMR and his daily physiology data from his PHR (Personal Health Record from an Apple, Google or Fitbit cloud- weight, blood pressure trajectory, blood sugar, diet, exercise, sleep pattern, stress levels). It would create a mind map summary synopsis of the patients history, identifying and marking important connections and parametric correlates from past history (adherence, medications, treatment plans, responses and outcomes) tying it to current situation as well as probable causes and potential treatment recommendations to his physician. Alan would not even have to travel to the clinic. He could have a Telehealth consult with his physician who would now be better informed and better armed.

Use Cases:

  • Precision medicine- the ability to customize a treatment plan based on a individual's physiology and disease characteristics

  • Moonshot initiatives to find cures for complex diseases such as cancer

  • In an outcome driven world where costs have to be optimized work will drive to the lowest cost producer - Physicians will work at the top of their license and common ailments will be triaged and diagnosed by Nurse Practitioners in retail settings. AI and Machine learning algorithms could then work as nurse assistants with patients.

  • The healthcare system even in these days relies on technologies such as fax to exchange information. Some of these are scanned and input into an EMR system. Physicians also enter diagnostic information as notes in a patient chart. Unfortunately the clinical and diagnostic taxonomy is not standard and some of the abbreviations used by clinicians could be interpreted differently by different clinicians. Natural Language Processing (NLP) can be used to parse these documents convert them to be editable, index them appropriately with semantic context to provide Machine Learning algorithms with data to diagnose more accurately.

  • No physician has time to go through patient records during a clinical visit. NLP can run through a patient history in the EMR and help build a one pane of glass view of the patient's clinical journey, arming a physician with a broader 360 degree view of the patient.

  • An AI system can create an informed decision framework modeling clinical pathways and proving available choices at each decision stage with relevant current literature (even to the extent of dumbing down medical terms into plain language) as well as a personal assistant for patients with complex diseases such as cancer

  • Clinical Logistics - An AI platform could be trained on signals from bio-medical devices (heart monitors etc.). It can predict the onset of an episode long before its occurrence. This can be used for managing conditions such as Sepsis for patients out of a surgical procedure in Acute or ICU settings. IBM's Watson with Medtronic can predict the onset of a hyper/hypoglycemic episode at least 3 hours before its onset. The Clinical logistics system could trigger or initiate a rapid response a lot quicker that observers who monitor the systems occasionally

  • Operational logistics- An AI platform can be deployed to improve clinical operations. Visual images in an OR can be used by an AI system to accurately estimate material usage for billing purposes. The systems can be deployed for better asset management and room availability, initial triage of patients in an ER and registration.

  • Virtual assistants with connectivity to patient information augment contact center agents shortening wait times for patients. Virtual assistants using AI/ML platforms can also discuss diagnostic results with patients relieving physicians and nurses to deal with that

  • A healthcare system on an average loses 3% of revenue through claims that are denied. Many of these are recoded and submitted. 67% of rejected claims are appealable. An AI system can analyze past history of rejected and resubmitted claims and assist with new claims- recoding appropriately and with risk adjustment based on patient history, reducing dramatically the number of denials

  • An AI system can analyze past patient demand and epidemics to predict patient volume to help deal with staff shortages.

  • NLP can help in charting at the bedside and converting the documents appropriately, helping with medication/diagnostic orders, alerting respective resources that need to act on this information- eliminating delays associated with batching as well as any open loop, passive process

  • An AI system can analyze incoming referrals, flag those that lack desired information, and route referrals to appropriate physicians, closing any communication loops using virtual assistants with the patient to set appointments or to gather any incomplete information

  • An AI system can advise a physician on the most appropriate drug/formulary based on patients history and current medications

  • An AI/ML system can help to analyze images effectively leaving a radiologist to focus on the most complicated ones

  • An AI/ML system can proactively look at bio medical equipment history and determine those that need calibration. It can also help with asset tracking and theft prevention

  • An AI/ML system can be trained on patients that were readmitted within 30 days of a procedure, using procedure done, socio economic data, zip codes, disease type, age and sex, co-morbidities, compliance characteristics and then be able to predict the likelihood of patients that could be readmitted. This would allow care coordinators/case managers to more effectively manage these patients

  • An AI/ML system can be used to monitor chronic patients at home and alert case managers to those patients that would need a high touch intervention

Some Issues

  • Debugging neural networks is a complex and hard problem. There are no standard methodologies to visualize the results. Algorithms need to be tested on training data and proven. That is a leap of faith for some- data lies siloed in a variety of incompatible systems in some environments (Multiple incompatible EMR's, PACS and RIS). This is a tough problem, but one that needs to be addressed anyway more so in an outcome centric healthcare system. Healthcare systems will have to break the silos, normalize the data and build training data sets as well as the data lakes for decision support.

  • Machine language algorithms build over a period of time. They operate on a trial and error framework where the training accuracy improves over time (deeper neural networks). Physicians lack patience with the algorithmic learning process , for inaccuracy in prediction leads to trust issues. AI/ML require change agents and are more likely to succeed in teaching hospital environments where there is greater tolerance for research and experimentation.

  • Management solutions for version control of algorithms as well as solutions to secure algorithms need to be provisioned

  • Marketplaces have to be built to securely distribute algorithms

  • Data scientists need access to data to develop and test algorithms. HIPAA and privacy requirements at time play spoilers in many a situation. What is the source of test data? Where is it staged? How is it secured? Data security is a big issue in healthcare.

The day may not be far off where Dr. Watson becomes the new Dr.House bringing much needed joy to Alan who face a more predictable, personalized and informed healthcare system.



Check back soon
Once posts are published, you’ll see them here.


  • Grey Facebook Icon
  • Grey Twitter Icon
  • Grey Instagram Icon
bottom of page