Human Intelligence, is the ability to adapt one’s behavior to fit new circumstances. In this article I discuss what would one need to build an AI-based solution.
In Psychology, human intelligence is not regarded as a single ability or cognitive process but rather as an “array” of separate components. Research in building AI-systems has focused on the following components of intelligence: 
These components of human intelligence are also utilized during diagnosing a patient and defining the treatment plan and protocol for the patient.
The process of Medical Diagnosis
The process of how a Doctor goes about her diagnosis of a patient, is the ability of a Doctor to adapt to varying presenting illnesses of her patients:
- Identify the main complain of a patient
- Gather information about the history of present illness
- List the possible diagnosis & record the differential diagnosis for a patient
- And then perform relevant diagnostic tests to determine the most likely causes for the presenting complaints
The doctor initiates the process of identifying the most likely cause of the patient’s presenting illness and then based on the results of the diagnostic tests, proceeds to confirm a diagnosis and then proceeds towards defining a treatment plan for enabling the patient to recover from the disease.
In the above simple process defined for a medical diagnosis, the Doctor (based on her training) makes use of all the “components of intelligence” to arrive at the most likely treatment plan for a patient. The process obviously gets more involved and complex depending on the type and nature of diagnosis.
Medical Diagnosis or Medical Algorithms?
From the above “very simple example” it’s clear that the doctor uses her learning and reasoning to proceed towards the best possible treatment pathway for the patient. And this can be treated as a series of Questions that help the doctor arrive at the “confirmed diagnosis” for the patient.
The process of Medical Diagnosis can then be treated as an Algorithm that helps the doctor arriving at a conclusion based on the presented facts.
Dictionary defines an “Algorithm” as, a process or set of rules to be followed in calculations or other problem-solving operations
The doctor in the above scenario has being processing via a set of rules and calculations and problem-solving operations to arrive at the confirmed diagnosis.
The doctor goes through a perception analysis to determine what specifically is presented based on the patient’s illness and then determines based on, not only the diagnostic test results, but also based on other parameters of a patient’s active and confirmed diagnosis.
Medical Diagnosis work in clinical practice generally has four models: 
- Pattern Recognition, wherein the doctor recognizes the current patient’s problem based on her past experiences with other patients, e.g., Down’s syndrome.
- Hypothetico-deductive, wherein the doctor performs a certain battery of tests to test a hypothesis, a tentative diagnosis.
- the Algorithm Strategy: the algorithm strategy has been used in Healthcare and has been represented using Medical Logic Modules , Arden Syntax for Medical Logic Systems  and Clinical Pathways  and finally the
- Complete History Strategy has been defined to be the identification of Diagnosis by possibility. Evidence based medicine is then used to come to a conclusion of the final diagnosis. 
The training process to arrive at a Medical Diagnosis has been used in the past to the development of expert systems or Clinical Decision Support Systems (CDSS). Early medical AI systems have tried to replicate the clinical training of a doctor into meaningful implementations of AI in healthcare.
Use Case for Artificial Intelligence in Healthcare
Understanding the process and workflow in healthcare is going to be important in implementing solutions that are “aware” and intelligent. And the systems that need to be developed for Healthcare need to be able to assist the clinicians with systems that are more close to the clinicians natural daily workflow.
Consider the current scenario of a physician meeting with a patient in a clinic setting, with the current systems in place the “Patient Visit” workflow generally involves the doctor having to divide her time between talking to the patient, examining the patient and recording the findings on an EHR (electronic health records) system. Most such visits can last from 5 minutes to an hour depending on the specialty (for instance, general medicine to mental health). Additional complexity is added to the workflow based on the patient diagnosis.
There have been many studies that have recorded the doctor’s reasons for resistance to enter the visit data into a system . A time and motion study of a patient – doctor interaction can be revealing in an EHR vs a non-EHR setting. While EHRs have shown their ability to reduce potential errors (as has been well documented in the report, to err is human) the additional steps of transcribing the visit data into an EHR is generally seen by the doctors as being a disruption in their natural visit or encounter workflow.
On the other hand, take into consideration a study of the workflow of a pathology department such as biochemistry or hematology, where the technology implementation is relatively easily accomplished. The pathology departments main “Entity”(from a systems perspective) to be processed is the patient sample and the level of automation required to process the various tests that need to be performed on the sample is quite well defined by its degrees of freedom, the test ordered by the doctor. Similarly the entity in a radiology department is the image that is the outcome of a radiology exam.
In radiology department for instance, an AI-based solution can enable operations at scale for enabling reading of radiology images from rural areas, where in the images get uploaded by the medical assistant or radiographer at the remote location. The AI systems now have the ability to read and report the images with increasing accuracy, but we still have some way to go before we achieve a greater deal of accuracy.
On the other hand, the “Entity” in a patient doctor interaction in a visit, the patient, has many more touch points within the patient care continuum and the level of complexity of this interaction needs to be dealt with in a completely different approach. While the processing in a pathology or radiology department is based on the sample or an exam, which is a snapshot at particular point in time, the treatment of a patient constantly needs to be monitored and presents more data points on an ongoing basis.
An AI-based solution to help a physician therefore needs to be applying for instance, the four models of medical diagnosis to a patient visit before we can call a patient visit as an intelligent or aware encounter.
If a doctor divides her time between listening to the patient regarding her present illness, and simultaneously recording the information on a computer system, there has been a disruption in the doctor’s natural workflow of focusing on the patient, of listening to her present illness, asking questions about onset, etc. and reviewing the results of the investigations and radiology reports. The doctor is trained to handle all these data points and process the information from the perspective of the four aspects of the medical diagnosis training of the physicians.
Here is an interesting story you would like to review showcasing a doctors 35-hr shift in Delhi, India. By the way the story lends itself to creating some really interesting “Intelligent Digital Assistants” for the doctors. It also presents to experts developing AI-based solutions for Healthcare, a fantastic time and motion study of a Doctors’ shift and the touch points to where the technology can be integrated into the Doctor’s “workflow”.
Current systems do not allow that, they tend to focus on implementing a strategy of recording by exception, by recording only the exceptions and all the other aspects being marked as normal, for instance. While such aspects have been proposed and devised by working with the physicians, still they are workarounds to do what the technology of today allows or allowed in the past. These are re-creations of paper based systems that have been translated to an electronic health recording system.
The Patient – Physician interaction needs to be revamped, in the current information technology systems by enabling the various components of human intelligence we have highlighted earlier:
- perception, and
An ideal scenario for a Patient – Physician interaction would be the implementation of a solution that “records” all of the conversation during a visit and automatically creates the Visit note, by understanding the Chief complaint, presenting illness, history of the patient, procedures ordered, medications prescribed, follow-ups or referrals ordered, et al. Purely based on the conversations between the doctor and the patient.
Such a scenario requires the implementation and collaboration between various components of the AI ecosystem. And that will be the true and useful implementation of AI for the Patient and Physician interaction, enabled by Artificial General Intelligence capabilities.
The change needs to be implemented by not only incorporating the changes to the core algorithms, but it also involves incorporating changes to the UI and UX design changes. AI based solutions will force a change in the way current systems have been designed. It is important to explain the way the physician thinks while interacting with the patient.
It’s been of late seen technology solutions to be hindering the doctor patient visit process. And hence it my endeavor to try to present the case that AI while hyped to be replacing doctors, is not yet ready for the prime time. There are areas of immense potential, radiology image processing for instance but then that’s from a process improvement perspective. And not doctor patient interaction perspective.
For years now, technology in healthcare has been trying to take the paperless approach and has tried to “replace” paper while forgetting that there is a more important component of enabling workflow in the Patient Care Continuum.
And it’s because of this reason, I argue that whilst it’s great for the technology hype cycle to see AI as the deliverer, we need to remind ourselves once again, that it’s not about going paperless, but ensuring the 15 min that a patient gets of the doctor’s time, are well spent with the conversation being patient focused and the technology receding to the background and generating the relevant care records.
In other areas of healthcare too it is about process improvement.
And add to that the fact that in most implementations in healthcare, clinical documentation is either cumbersome or non-existent, the hype cycle of AI needs to consider these issues. From my understanding since the underlying data is fragmented, not standardized and not interoperable in majority of the instances; I took a shorter term view of the AI implementation in the systems in this article.
Current Status of Artificial Intelligence in Healthcare
There has been data explosion in Healthcare not only from the perspective of the patient care continuum, but also from the point of view of the resource management and scheduling, inventory and purchase management, insurance, financial management, etc.
While most of the current focus has been on building AI-based solutions that are in the patient care continuum, there are definitely many more areas within a healthcare organization that will benefit from the implementation of intelligent systems.
Just the other day, I attended a conference around AI and the panelists were mentioning the following uses of AI:
- ecommerce recommendations
- learning for students based on concepts in school
- autonomous cars
- AI-based treatments plans for cancer patients
- intelligent assistants, chatbots
- teaching computers to see; etc.
And while they all highlighted areas of advancement in AI tech, they are yet to reach the ability to currently create a system that converts a doctor patient conversation to actionable events that can spawn workflows that needs to be instantiated based on the ever changing patient condition.
In the near-term, I see there will be specialized implementations of AI that will enable the brute power of technology to present the best case scenario for a particular patient condition, but an AI Physician is still a work in progress. This has been shown to be a success with the advent of cancer care solutions using IBM Watson.
The AI systems are being implemented in various scenarios in healthcare and you could consider them to being “trained” and being presented with a great amount of data and studies. As more data is presented to these AI systems, their level of accuracy will only improve and provide benefits in-terms of scale and reach thereby reducing the time to diagnosis and time to treatment for patients having affordability and accessibility issues in healthcare.
Artificial Intelligence has already started making its way into healthcare, with 90+ AI startups getting funding to deliver solutions like:
- helping the oncologist define the best treatment plan specific to each patient
- a virtual nursing assistants, to follow-up with patients post discharge
- drug discovery platforms, for new therapies
- Medical Imaging and diagnostics
- The use of AI in diagnosing diseases, patient education and reducing hospital costs
- You can also find a great discussion on machine learning, wherein how machine learning could replace/ augment doctors via the health standards podcast with Fred Trotter.
Some of the other areas where AI is being implemented in Healthcare. Microsoft, Apple, IBM and other major players are all looking to AI help in curing people. And they are forming a group that creates the standard of ethics for the development of AI.
AI in healthcare also has a potential to be leveraged to be implemented in the following aspects of Healthcare Industry:
- Billing and Insurance Workflow, Insurance reconciliations and provider workflows can be enhanced by enabling total automation of the processes by enabling handling of the insurance claims by AI based Insurance agents. The exceptions and outliers can be escalated for manual interventions and closures.
- Improving customer experience in healthcare by providing a 360 degree engagement, the SMAC based solutions will use the power of integrating the data streams from multiple sources to help deliver a better service to the patients.
- Inventory and Supply chain processes can benefit from AI driven optimization by incorporating e-commerce driven innovations that allow for a democratization of product to vendor mix by searching and delivering the best cost options to the procurement department. Thereby bringing the costs down. Logistics improvements delivered in other industries need to come to healthcare to allow for the reduction in the cost of procurement of drugs, devices and durables. AI will help organizations in identifying variable costs and help them understand how to handle scenarios that will present themselves in an ongoing basis.
- AI enabled resource management and scheduling will allow for identifying areas that need to be staffed with more resources and when additional resources need to be hired to meet with the increasing demands or provide elastic resource management based on ever changing operational demands. Booking appointments with doctors, will become a job taken up by Bots or AI assistants, enabling the nursing and administrative staff to focus more on delivering care and enhanced service experience for the patients.
- AI-based people management systems will help hospitals in recruitment, retention and performance management of their employees. By presenting an analytics driven approach to people management, systems will be able to help employees to be trained to take up newer roles and responsibilities.
So by when will AI really take over Doctors?
It’s clear from the image above, that estimates of how much processing power is needed to emulate a human brain at various levels (from Ray Kurzweil, and Anders Sandberg and Nick Bostrom), “along with the fastest supercomputer from TOP500mapped by year. Note the logarithmic scale and exponential trendline, which assumes the computational capacity doubles every 1.1 years” . Kurzweil believes that mind uploading will be possible at neural simulation, while the Sandberg, Bostrom report is less certain about where consciousness arises
Based on the above point of view, an interesting question to ask today:
If a Doctor goes through 7+ years of training to become a specialist, how many days will it take for an AI-based Physician?
The answer perhaps lies in the following statements:
Chief scientist and AI guru Andrew Ng of Chinese search giant Baidu Inc.once put it,
Worrying about takeover by some kind of intelligent, autonomous, evil AI is about as rational as worrying about overpopulation on Mars. , .
What is it that makes us human? It’s not something that you can program. You can’t put it into a chip. It’s’ the strength of the human heart. The difference between us and machines.
– Terminator Salvation, 2009
: AlanTuring.net What is AI? http://www.alanturing.net/turing_archive/pages/reference%20articles/what%20is%20ai.html
: Viewpoint: Data-driven, interaction biases prove challenges for AIhttp://www.beckershospitalreview.com/healthcare-information-technology/viewpoint-data-driven-interaction-biases-prove-challenges-for-ai.html
: Improving Diagnosis in Health Care | The National Academies Press https://www.nap.edu/catalog/21794/improving-diagnosis-in-health-care
: The diagnostic process in general practice: has it a two-phase structure http://fampra.oxfordjournals.org/content/18/3/243.full
: Managing Medical Logic Modules.https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2247534/
: HL7 Standards Product Brief – Arden Syntax v2.9 (Health Level Seven Arden Syntax for Medical Logic Systems, Version 2.9) http://www.hl7.org/implement/standards/product_brief.cfm?product_id=290
: Clinical Pathways via Open Clinical, knowledge management for medical care http://www.openclinical.org/clinicalpathways.html
: Sackett DL, Haynes RB, Guyatt GH, Tugwell P. Clinical Epidemiology. Boston: Little, Brown and Co., 1991; 3–18.
: Barriers for Adopting Electronic Health Records (EHRs) by Physicians https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3766548/
: Artificial General Intelligence. https://en.wikipedia.org/wiki/Artificial_general_intelligence
: AI guru Ng: Fearing a rise of killer robots is like worrying about overpopulation on Mars http://www.theregister.co.uk/2015/03/19/andrew_ng_baidu_ai/
: The Artificially Intelligent Doctor Will Hear You Now. https://www.technologyreview.com/s/600868/the-artificially-intelligent-doctor-will-hear-you-now/
: Why we are still light years away from full artificial intelligence |https://techcrunch.com/2016/12/14/why-we-are-still-light-years-away-from-full-artificial-intelligence/
: AI In Healthcare Heatmap: From Diagnostics To Drug Discovery Startups, The Category Heats Up https://www.cbinsights.com/blog/artificial-intelligence-healthcare-investment-heatmap/
: Doctor’s 35-hr shift on 8 bananas, a toilet in nearby cafe http://indianexpress.com/article/india/india-others/doctors-35-hr-shift-on-8-bananas-a-toilet-in-nearby-cafe/
: Gigerenzer’s simple rules by NS Ramnath on Founding Fuel http://www.foundingfuel.com/article/gigerenzers-simple-rules/
: A.I. VERSUS M.D: What happens when diagnosis is automated? By Siddhartha Mukherjee http://www.newyorker.com/magazine/2017/04/03/ai-versus-md