I & L to #AI & #ML in Healthcare by Jyoti Sahai, @jyotisahai

Have you ever wondered why if confronted with any illness symptoms that appear even a bit abnormal, we prefer to consult with a doctor in a large hospital only, even though a more competent doctor may have a clinic next door itself.

And have you ever wondered that what that preference has to do with Artificial Intelligence (AI) and Machine Learning (ML)!
To explain that, let me recount what happened to me twenty-one years back. I vividly remember that incident from 1997 that I can now relate well to the significance AI and ML are having in healthcare currently!

SYMPTOM

I was being examined by a leading physician at Agra (who had an experience of over twenty-five years and had a roaring practice) for a pain near my left toe. The conversation progressed as follows:
I: Doc Sahib, please have a look at my left toe. I am troubled by a severe pain for over last three weeks. I cannot put my foot down or wear the shoes even.
Doctor: Did you ever have this type of pain before?
I: No
Doctor: (Examining the pain area closely) Do you feel any irritation, or feel any urge to scratch that area?
I: No.
Doctor: Do you eat lot of red meat?
I: No! I am a vegetarian.
Doctor: Do you like to eat lot of tomatoes, or cheese, or spinach or any other high protein foods.
I: Yes. Very frequently have cheese-spread, and baked beans in breakfast, and of course tomato in some form is generally there in all meals.
Doctor: (Prepares a slip for the diagnostic lab) Please have the uric acid blood test done as I suspect you have gout.
I: Thanks, Doctor. Will come back later with the test results.
(Later during the day)
I: (Handing over the lab report) Here Doc Sahib. Please have a look at the report.
Doctor: (Going through the test report) That is what I thought. You have gout! Your uric acid level is 12.4 mg/dl which ideally should have been between 3.5 mg/dl - 7.0 mg/dl. I will immediately start the medication.
(The doctor then spent few minutes to explain what gout was and how it impacted my health, and my lifestyle.)
I: Any restrictions on diet?
Doctor: Yes. For the time being completely stop eating your favorites - cheese, tomato, spinach and all dals (lintels) except 'moong' dal.
The treatment started that same day, and within three weeks the pain had substantially subsided, and gout was well under control.
What I have narrated above was actually the Step 3 of the treatment plan that I had followed for almost three weeks before I met that doctor at Agra.
  • The symptoms - After having spent more than five years in Bangalore I had just moved to Noida and had started to adjust to a different living (and professional) environment. One day I woke up to acute pain in the area near my left toe. It appeared a little swollen and made it difficult for me to even wear the shoes.
  • Treatment Step 0 - As usually happens with all of us, initially I tried out some home remedies only, like soaking the leg in warm water and taking some pain killers. That was to no avail and the pain persisted.
  • Treatment Step 1 - A few days later I had to attend a family gathering where a relative of mine, fresh out of college after completing her course in medicine, had a look at it and opined that it could be some allergic reaction due to change of location (from Bangalore to Noida) and prescribed some tablets. However, the pain still persisted and even increased after few days of that treatment.
  • Treatment Step 2 - It was then that I decided to consult a practicing physician and went to a clinic just across the road where we lived. He examined the pain area and diagnosed it as some sort of inflammation and advised putting poultice for few days. Even after several days of that treatment, the pain did not subside but actually aggravated.
  • Treatment Step 3 - Experiencing no relief for over three weeks, I finally decided to consult my younger brother, a leading plastic surgeon at Agra, who took me to one of his colleagues who was a leading physician. What happened next, I have already stated above.

DIAGNOSIS

Now after twenty-one years when I analyze that line of treatment, I realize that
  1. The young doctor who first advised me maybe had never seen such symptoms earlier and thus was not able to diagnose correctly.
  2. The physician I consulted next might have seen only a few patients with a similar set of symptoms that I had (but not with the same illness), therefore was not able to formulate the right questions to ask that could have led to the correct diagnosis from a set of possible outcomes arising from similar symptoms.
  3. However, the doctor at Agra with his vast experience, had obviously seen those set of symptoms several times earlier and had acquired sufficient I and L to treat such cases effectively.
  4. Thus, though all the three doctors were surely competent, what the first two doctors obviously lacked were
  • the ability to apply their I (Intelligence) in (a) arriving at the correct diagnosis based on the symptoms they were presented with, and (b) subsequently determining an appropriate line of treatment; and
  • the extent of L (Learning) that comes with experience of treating hundreds and thousands of patients with various types of symptoms possible that brings in the knowledge that what could be the possible diagnoses and what treatments worked or did not, and why or why not?

LINE OF TREATMENT

By applying AI and ML techniques and solutions in healthcare it may now become possible to make available the accumulated I and L - resulting from the large number of successful (and unsuccessful) treatments by various experienced doctors - to those competent but less experienced physicians.
With access to an appropriate AI/ML system, even a physician in a small clinic in a remote location could
  • draw upon the accumulated experience of other successful doctors;
  • be guided properly to arrive at the correct diagnosis and subsequently to determine an appropriate line of treatment; and
  • confirm that the planned line of treatment is suitable for the medical profile of the patient. In case the patient's medical profile is not readily available (like in case of emergency patients or admittances to trauma centers), AI/ML systems could caution the first medical responders on the possible complications (if any) associated with any planned line of treatment.

OUTCOME

Effective use of AI/ML systems in healthcare can deliver sustained benefits for all relevant stakeholders:
For the patient:
  • Assurance that the physician would arrive at a correct diagnosis, and would propose an appropriate and effective line of treatment with less or almost no margin of error;
  • Obviating the need to rush to a larger hospital/clinic just because the symptoms are a bit abnormal; and
  • Faster and more effective response from medics in emergency cases.
For the healthcare provider:
  • Increased efficiency with lower turnaround time for patients;
  • Faster and accurate diagnosis and effective treatment;
  • Substantial reduction in unfair treatment cases; and
  • Substantially faster and accurate response by first medical responders in emergency cases.

FOLLOW-UP

In the face of massive disruption taking place in healthcare space, and the frantic pace of medical data generation, any AI/ML system is likely to be soon become outdated, ineffective and irrelevant, if it is not constantly updating its Intelligence and is not constantly Learning.
Thus, it is imperative that all instances of successes and failures, arising out of using any AI/ML system, are fed back into that system to ensure constant refinement of its algorithms. That will result in it providing even more accurate outcomes for future users.

Conclusion

From the above it is evident that an AI/ML system can be a powerful ally of a physician and its deployment should not be termed as "man against machine" by any means.
In my opinion, AI/ML technologies are still meant to assist the medical fraternity and are not really likely to replace doctors (at least in foreseeable future)!
The article has been republished here with the authors permission. The article was first published in the authors' linkedin pulse page.

Author
Jyoti Sahai
Chairman and Managing Director at Kavaii Business Analytics India Pvt. Ltd. Jyoti Sahai has over 42 years of experience in banking and IT industry, and is currently the CMD of Kavaii Business Analytics India. Kavaii provides analytic solutions in Healthcare and IT Services domains.
Read more »

Natural Language Processing #NLP - Giving doctors the freedom to write what they want by Dr. Anuradha Monga

Healthcare produces the highest quantity of data records as compared to any other industry. There has been a substantive shift in the provider workflows from capturing data in paper based records to electronic modes and storage in the past few decades.



Natural Language Processing: Giving doctors the freedom to write what they want

Electronic health records (EHRs) have clearly emerged as an innovative technology to facilitate the transition. However despite of the advancements, EHRs have not been able to achieve credible benefits in areas of population health management, health information exchange, patient care coordination and clinical analytics. 

One of the biggest barriers in achieving success with EHRs has been the disparate forms of data which are difficult to aggregate and analyze. Doctors feel comfortable writing notes along with the flow of their clinical thoughts, however EHRs are not designed to capture medical information in a doctor’s natural language. This inability many a times leads to poor EHR usability. As a result, a lot of valuable information is left out from the ambit of analysis. With the advent of newer technologies, now it may be possible to plug such gaps. NLP (natural language processing) is one such technology which providers are now adopting with an anticipation to improve clinical outcomes and for the simplification of the daunting task of data entry in a computer.

Clinical data is not consistent, making analysis difficult

An EHR captures data in primarily four ways:
  • Clinical Data is directly entered in pre-structured templates 
  • Scanned documents are uploaded in the system 
  • Text reports are transcribed by speech recognition technology or by dictation and manual data entry.
  • Data is purged into an EHR by interfacing it with other information systems like laboratory systems, radiology systems, or monitoring devices. 

Clinical data is usually presented in a structured or unstructured format. Selective choices for capturing data in the form of templates like physician order sets, drop down menus, check boxes etc constitute structured data. Aggregation, analysis and reporting from structured data is easier but doesn’t provide an individualized, customized identity to an EHR. On the other hand, unstructured data constitutes free text narratives and clinical notes i.e doctor’s notes, patient encounters, patient health records etc and enable the physicians and patients to get their observations, complaints and concepts recorded in their own parlance. The unstructured data is a rich source of information about a patient’s health but it’s a challenge to transform it into structured and analyzable data that can be used for improving care outcomes. This challenge can be overcome with the technology of natural language processing.

Unstructured clinical notes are a mine of golden data; the wait to explore them ends with NLP


NLP is a data science based technology that can extract data from free text. NLP can be used by clinicians to convert medical notes into formats which are structured and standardized. Auto-processing of textual data can help providers in making use of clinical documentation data for a variety of purposes including but not limited to:


  • Improving communications between healthcare teams and thus help improve outcomes
  • Reduce overhead costs of clinical documentation
  • Improve revenues by automation of the coding and documentation

Computers can be given the ability to infer the intended meaning of words, thus enabling them to identify trends and patterns in huge datasets. 

NLP can change the course of the way chronic diseases are managed:

One of the most promising area for exploring use cases of NLP in healthcare includes predictive analytics and risk scoring. Carefully deployed AI tools can be used for risk stratification and determination of hotspots in chronic diseases. 

NLP can be used to tag socioeconomic terms hidden in free text notes to identify the social determinants of health. This can be augmented with machine learning to develop risk scores by proactive identification of trends from clinical and social data, laboratory reports, diagnoses etc. It is possible to create algorithms and train them on clinical record data to identify disease symptoms accurately. 

Clinical records are a rich source of information regarding the symptoms of many diseases. Grouping of such similar symptoms can help in syndrome identification on the basis of disease presentation. As a result, it may be possible to unearth clusters which may otherwise not be suspected. Routinely available information in electronic health records, such as demographic and geographical location data and primary care free-text clinical records should be leveraged while making use of such algorithms.  

Why off the shelf NLP engines may not be what the doctors want:


While it sounds easy, healthcare free text data comes with its own challenges. Word sense ambiguity is perhaps one of the most challenging problems in the noise of free text clinical notes. Accurate translation of the structured patient information pertaining to medical procedures, symptoms, tests etc depends on the algorithm’s ability to assign correct interpretations to the relevant medical words. For example, the acronym RA can be used in different contexts with different meaning by doctors. RA can be interpreted as right atrium, right arm or rheumatoid arthritis depending on the case presentation and clinical context. 

Disambiguating the senses of acronyms, symbols and words that are used in a doctor’s clinical notes can significantly ease the burden on human effort needed to develop more accurate systems. A data-driven approach which involves development of any algorithm that infers patterns should consist of a supervised and unsupervised learning phase to yield benefits. In supervised learning every data item of the training data is labeled with the correct answer. Unsupervised learning on the other hand is a process where the computer recognizes patterns automatically. The true potential of an NLP and machine learning algorithm can only be harnessed when the data is trained in the provider’s environment.

Word sense disambiguation based NLP pays a significant role in improved analytics and patient outcomes:


Word sense ambiguation based language processing ability of the computer for accurate mining of clinical documents can bridge the gaps in documentation and aid clinical decision support and clinical documentation improvement programs. 

More insightful extraction of data is possible with a decreased ambiguity in clinical data. When the computer has the ability to infer the intended meaning of words, it can find useful patterns in heaps of data easily. IBM’s Watson Supercomputer technology is an apt example of how NLP can facilitate meaningful analytics, by identifying such patterns. IBM’s content analytics process is used for collection and analysis of structured and unstructured data, and its similarity analytics makes use of NLP and machine learning technology for analysis of a large number of variables in a patient’s medical history and present condition to identify patterns and draw a comparison with similar conditions and potential outcomes. 

There is no doubt that word sense disambiguation enabled NLP technology can have a potentially huge on impact clinical data analytics with its superior ability to infer meanings of extracted data more accurately. Data analytics for improved patient outcomes is not the only benefit of this technology, it can also support accuracy of billing. With its ability to support clinical documentation improvement programs, it can also help in improving clinical workflows.

SymptomAI by “PredictDisease” is a healthcare analytics platform that is driven by artificial intelligence, NLP and machine learning to assist patients and primary care physicians by measuring the potential risk of a chronic disease that starts with minor symptoms. The platform leverages data from lifestyle activities, social media/website forums, scientific research papers, and family history, matching these with known signs/symptoms and other demographic characteristics for the early detection of the chronic disease. It takes into account, social and biologic determinants of health to predict the risk score. Visit us at www.predictdisease.com or write to us at [email protected] for more info.

References: 
[1]. Auto Coding and NLP: 
http://www.himss.org/content/files/AutoCodingandNaturalLanguageProcessing(WhitePaper).pdf

[2]: Dooling, Julie A. “Advancing Technology Connects Transcription and Coding: The Developing Role of NLP, NLU, and CAC in HIM.” Journal of AHIMA 83, no.7 (July 2012): 52-53

[3]: Goldberg, Michael. “IBM Makes New Health Care Push with Predictive Analytics, Process Management.” Data Informed. http://data-informed.com/ibm-makes-new-health-care-push-with-predictive-analytics-process-management/


Author
Dr. (Maj) Anuradha Monga
A versatile military veteran with expertise in healthcare management, Anuradha has acquired real world experience in areas of Hospital operations, Health insurance claims management and mass insurance, Healthcare IT, NABH implementation and digital marketing.
Read more »



POPULAR POSTS

Popular Posts