Lab Soft News

AI Platform to Assist Hospital Physicians in Lab Test Ordering

I have blogged extensively about the deployment of AI in healthcare (see, for example: CMS Offers $1M Reward for AI to Predict Unplanned Hospitalizations). I have learned that the simplest AI applications may sometimes have the largest beneficial effects (see: Medical Community Braces for Algorithm to Reduce Unnecessary Imaging Orders). That was my reaction when I read a recent Dark Daily article (see: Florida Hospital Utilizes Machine Learning Artificial Intelligence Platform to Reduce Clinical Variation in Its Healthcare...). Below is an excerpt from it:

Variation in how individual physicians order, interpret, and act upon clinical laboratory test results is regularly shown by studies in peer-reviewed medical journals to be one reason why some patients get great outcomes and other patients get less-than-desirable outcomes....At Flagler Hospital....,a new tool is being used to address variability in clinical care. It is a machine learning platform called Symphony AyasdiAI for Clinical Variation Management (AyasdiAI) from Silicon Valley-based SymphonyAI Group. Flagler is using this system to develop its own clinical order set built from clinical data contained within the hospital’s electronic health record (EHR) and financial systems....

Clinical order sets are designed to be used as part of clinical decision support systems (CDSS) installed by hospitals for physicians to standardize care and support sound clinical decision making and patient safety. However, when doctors don’t adhere to those pre-defined standards, the results can be disadvantageous, ranging from unnecessary services and tests being performed to preventable complications for patients, which may increase treatment costs....Flagler first used the AI program to examine trends among their pneumonia patients....[A] “smart software solution” will be inserted into the clinical workflow of physicians. This solution will automatically guide the physician to follow the established care protocol. In turn, this will give the medical laboratory the simple role of accepting a lab test order, performing the analysis, and reporting the results.

My initial reaction to the use of lab order sets years ago was that they were too rigid and bureaucratic. Having said this, I did realize, even early on, that there was a place for them. Many physician continue to use lab testing strategies that they learned years before as residents. Some relevant and necessary tests may thus be overlooked for admission lab test orders. However, I was struck by the following sentence in the excerpt above: Clinical order sets are designed to be used as part of clinical decision support systems (CDSS) installed by hospitals for physicians to standardize care and support sound clinical decision making and patient safety. 

We now have the ability to develop test order sets that are more flexible and sophisticated and integrated with clinical and lab test data in the EHR and supported by predictive analytic tools (see: Diagnostic and Predictive Analytics and Their Possible Link to the Future of the LIS; Three Examples of Predictive Analytics in Clinical Practice). The deployment of sophisticated CDSS thus has the potential to diagnose patients faster as well as reduce length-of-stay (see: Mayo Develops CareSelect Tool to Guide Best Lab Practices for Clinicians; Dr. Alexa Will See You now -- An Extension of Paging Dr. Google).

Big Tech Initiatives in Healthcare Coming from Many Directions

I have posted a number of notes about the increasing role of Big Tech in healthcare. It's time to take a look at the big picture and helpful in this regard is a SWOT table that was included in a recent article about this topic (see: BIG TECH IN HEALTHCARE). It was compiled by Business Insider Intelligence and is presented below:

After having reviewed the bulleted items in the four columns for Alphabet (Google), Amazon, Apple, and Microsoft, I was impressed by the breadth and complexity of the list. For me, all of this this distills into the following Big Tech initiatives in healthcare:

• Home monitoring with wearables for wellness and diagnosis (see, for example: The Evolution of "Wearable Health Ecosystems" and Associated Partnerships)..
• AI and data analytics (see, for example:Three Examples of Predictive Analytics in Clinical Practice; Analysis of the Failure of IBM's Watson; Vertical Integration Backfires).
• Natural language processing of EHR records (see, for example: Amazon Launches New Medical Record Language Processing Service).
• Retail prescription drug sales (see, for example: Amazon-PillPack Pushes Against CVS with Prescription Deliveries; Also Seeking PBM Status). 
• Data collection for wellness research (see, for example: Google Said to Be Seeking Data and B-to-B Relationships with Fitbit Purchase; Apple Getting Traction with Its App to Access EHR Records with iPhones).
• Software to increase efficiency of hospital operations (see for example: Dictating EHR Physician Notes Using an Apple Watch with AI Assistance).

Here's are my general conclusions about the role of Big Tech going forward :

• Big tech is banking on healthcare for a large contribution to its future earnings (see: Healthcare Will Contribute a Sizable Portion of Future Earnings for Apple).
• The Big Tech focus will be on AI, home monitoring and diagnosis with wearables, hospital and home data collection for research, and prescription drug sales. 
• Home health monitoring systems and predictive analytics web sites from Big Tech may eventually account for a growing percentage of referrals to heath systems for disease treatment (see: Status Reports and Hospital Referrals from Wearable Health Ecosystems; Why Some Hospital CEOs May Be Opposing Interoperability in Support of Epic).

Analysis of the Failure of IBM's Watson; Vertical Integration Backfires

I have blogged periodically about IBM's Watson which has never lived up to its early expectations (see: Scandal at M.D. Anderson -- Operating Loss and Then Watson Deep-Sixed). Forbes recently published a piece about Watson with an analysis of this failure (see: IBM Watson And The Value Of Open) and below is an excerpt from it:

...[W]ith the passage of more time, it must be said that IBM Watson has not delivered the results that IBM expected. One particular moment was the decision of MD Anderson’s Cancer Center to withdraw from its partnership with IBM in 2017. An internal audit by the University of Texas found that the university had spent over $62 million dollars (not counting internal staff time) and did not meet its goals....While many factors were no doubt at work here, the shift in IBM’s focus from cognitive computing to cloud computing is certainly a part of this change....

There are technical reasons [that IBM has not succeeded with Watson including]...the rapid and unexpected progress of other forms of AI, such as deep learning. But another likely reason has to do with how IBM took its then-leading technology to market. IBM commercialized Watson in health care with a vertically integrated strategy. All of the activities from the laboratory to the customer were delivered by IBM itself. There were no third party system integrators. There were no APIs for others to use to link to and build upon Watson. There was no software development kit, or open source reference designs to help people explore Watson’s abilities, and figure out whether and how they might solve other problems. Even customers were only able to access a black box, putting queries in and receiving the result, with no way to comprehend how the result was developed. Being vertically integrated gave IBM complete end-to-end control over Watson. But it condemned Watson to being applied in only a few areas.

This story calls to mind my recent note about how Cerner was turning to AWS to solve its EHR functionality problems (see: Cerner Offers AI Tool to Combat Physician EHR Burnout; Cerner Leveraging Amazon's Cloud Storage and AI for Increased EHR Functionality; EHR Vendor Cerner Turns to Amazon Web Services for AI Development). Epic will undoubtedly pursue the same course of action in the future or face serious consequences. Here is the logic that I think is operating in both cases:

1. EHRs like Cerner and Epic are EHR giants with systems designed primarily to promote efficient billing. Inefficiencies in these systems are leading to physician and nurse burnout (see: Study Shows Physician Burnout Directly Related to EHRs). These highly trained professionals are being turned into clericals forced to use poorly designed systems. 
2. There is only one solution open to these companies to revamp their products -- AI. The inefficient antiquated software and inefficient design will be pushed out of sight and users will interact mainly with the AI-driven features.
3. I suspect that neither Cerner nor Epic have the employee talent to develop these AI add-ons despite their large size and profits in this burgeoning field. They are thus turning to third-parties experts like AWS to develop new AI solutions.
4. Users in hospitals will soon develop greater expertise in using the AI-enhanced products than the EHR employees because they will be interacting with them on a constant basis. They will then choose to interact more with the third-party developers than Epic and Cerner employees because that's where the correct answers will lie.
5. Similar to the IBM/Watson failure, Cerner and Epic are entering a cul-de-sac and will slowly loose control and support for their base products. The third-party AI companies will come to understand this and soon bring to market their own EHRs that will be modern and comprehensive. Ah, the perils of too much emphasis on vertical integration.

Why Some Hospital CEOs May Be Opposing Interoperability in Support of Epic

In a recent post, I discussed the opposition to the proposed HHS rule on EHR interoperability on the part of Judith Faulkner, founder and CEO of Epic, and the CEOs of some 60 health systems (see: Faulkner Opposes Proposed HHS Hospital Interoperability Rules; No One Surprised). In this note, I would like to discuss some possible reasons why the CEOs of health systems might oppose such a rule.

• Firstly and as I have long maintained (see: NYT Op-Ed on EHR Interoperability Blames Vendors and Greedy Hospitals), data portability enables healthcare consumers to "shop" hospitals. However, they may not always do so on a voluntary basis. Yearly churn in health plan membership is estimated to be about 25% so some consumers may be forced out of their current health plan due to, for example, a job change (see: Reducing Churn to Increase Value in Health Care: Solutions for Payers, Providers, and Policymakers). The bottom line here is that portability of EHR records is a key element in the improvement of consumer's healthcare experience.
• Secondly, it is asserted by hospitals that interoperability provides health insurance companies an unfair advantage over health systems in negotiating  reimbursement rates between the two parties. I believe, as do many others, that the entire hospital billing system is broken and hospital bills are often indecipherable (see: Two Reasons Why Medical and Hospital Bills Are Such a Mess). The whole hospital reimbursement system needs to be scrapped, allowing us to to start over and create a rational new system. Here's an article from 2006 that discusses this mess (see: The Precarious Pricing System For Hospital Services). I am thus in favor of any change that may disrupt the current hospital billing systems.
• Finally, I was struck by the point raised by Jonathan Bush who favors interoperability as referred to in my recent note (see: Faulkner Opposes Proposed HHS Hospital Interoperability Rules; No One Surprised). He said the following: "... with new data-sharing opportunities, patients can send their medical data to new tech services that can "aggregate data and recognize patterns in a way that health systems can't..." This is one facet of predictive analytics and an exciting new development in healthcare. I don't want to stifle its development in any way. I can envision that consumers will use such "tech services" as a check of a current diagnosis, particularly a serious one.

Virtual Personal Assistants as a Key Factor in Wearable Health Ecosystems

I have posted a number of notes pertaining to wearables and wearable health ecosystems (see, for example: Sony Develops Health Wearable Providing Connectivity, Stipulating That It's Not a Smartwatch; The Evolution of "Wearable Health Ecosystems" and Associated Partnerships; Status Reports and Hospital Referrals from Wearable Health Ecosystems). I have not yet raised the issue of providing a virtual personal assistant (VPS) with these home-based health systems. This idea was discussed in a recent article (see: Wearables Hold the Key to Connected Health Monitoring) and below is an excerpt from it:

As voice interaction with virtual personal assistants (VPAs) becomes standard on consumer devices and wearables, their functionality will become more personalized, opening up new opportunities for healthcare services. Some services will come straight from the next generation of AI features and could include alerts to health professionals or designated caregivers, as well as predictive algorithms to help prevent or manage lifestyle diseases such as type 2 diabetes and hypertension. Gartner estimates that by 2020, no-touch user interfaces (UIs) will account for about 15% of personal devices, wearables and Internet of Things (IoT) endpoints. A more natural product interface delivered with voice, movement and gesture will increase ease of use and help bring devices into mainstream healthcare use cases.

I had not heard the term no-touch user interface as a synonym for a virtual personal assistant but it makes perfect sense. Consumers are growing increasingly comfortable with the use of virtual assistants in their homes using technology like Amazon Echo Dots. Such virtual assistants will obviously be integrated into our home health technology  (see: The Evolution of "Wearable Health Ecosystems" and Associated Partnerships). They can remind us in a non-threatening way when we need to take certain health measurements such as blood pressure or take our prescribed medications (see: Cigna Unveils App to Monitor and Improve Medication Adherence).

Under dire circumstances such as a fall or extreme change in health status, the virtual assistant can also call for outside help. Virtual assistants can also be used to scan the web for suitable medical advice when queried. I have referred to this as "paging Dr. Google" or "Dr. Alexa" in older notes (see: Dr. Alexa Will See You now -- An Extension of Paging Dr. Google). Virtual assistants can also be programmed to only seek information from most reliable web sites and will thus improve the quality of the information being sought.

Amazon and AWS Help Improve Hospital Efficiency and Reduce Costs

In a recent note, I discussed how Cerner was turning to Amazon/AWS for help in providing assistance with AI technology that an be embedded in its EHR product. One example was the deployment of a chatbot that can be used to interrogate the EHR for critical information that can be difficult to extract, particularly in the form of natural language (see: EHR Vendor Cerner Turns to Amazon Web Services for AI Development). In a recent article about Amazon's healthcare strategy for 2020 (see: 15 things to know about Amazon's healthcare strategy heading into 2020), I was reminded of the relationship between Amazon and Beth Israel Deaconess Medical Center. Here is a brief quote from that article:

In March, [2019], Amazon made a $2 million investment in Boston-based Beth Israel Deaconess Medical Center to test artificial intelligence tools. The two have partnered since 2016 when Beth Israel bought AWS's cloud software and reported significant improvements with operating efficiency.

Here is a quote from another article about the same topic that provides more details about projects that have been initiated in this same program (see: Improving Patient Care with Machine Learning At Beth Israel Deaconess Medical Center)

An initial BIDMC research project used machine learning to optimize the schedules of its 41 operating rooms and align those schedules to improve patient flow in the inpatient setting. Another project has leveraged machine learning to improve operational flow within operating rooms. Now, incoming pre-surgical document packages will be scanned as images and processed with TensorFlow on Amazon SageMaker, hosted in BIDMC’s secure AWS cloud. This machine learning driven process automatically recognizes and inserts consent forms into corresponding electronic health records (EHR), saving hospital staff hours of manual work. BIDMC built a model that scans EHRs to look for key elements like a completed consent form. If a consent form isn’t found, a signal appears on the EHR and nurses follow up with those patients.

The median operating margin for U.S. health systems in 2018 was only 1.8% (see: Operating margins stabilize, but not-for-profit hospitals still vulnerable). Part of the reason for this situation is the nature of our entire health system with its eccentricities and perverse incentives. However, I will also fault some health system executives for these financial challenges. They have tended to focus over the years on optimizing revenue by negotiating with payers rather than vigorously trying to reduce their own cost structures. I am not sure at this point that the use of AI to increase the efficiency of health system operations will have a broad effect. However, this is certainly a trend to watch.

EHR Vendor Cerner Turns to Amazon Web Services for AI Development

We have reached the point when one of our major EHR vendors, Cerner, has come to understand that it can no longer rely on its own internal resources for AI and predictive analytics development (see: Cerner builds machine learning platform on Amazon Web Services). AI has become so sophisticated that EHR vendors like Cerner can't hope to compete in the field without support from experts such as those working for AWS. An excerpt from this article is presented below: 

In July...[Cerner] named AWS its preferred cloud provider. Under the agreement, Cerner is transitioning its artificial intelligence and machine learning workloads to AWS to develop predictive technology. The new platform, called Cerner Machine Learning Ecosystem, was built using several Amazon services and includes machine learning services Amazon Transcribe and Amazon Translate, among others. With the platform, developers will be able to create chatbots that give patients access to their personal medical records as well as the ability to ask questions about their medication, diagnoses and medical conditions. Cerner is also developing voice-activated documentation tools; the EHR vendor recently achieved its 500th patent with a new voice-assisted technology that supports clinician documentation.

I have blogged extensively about AI technology developments in diagnostic medicine such as predictive analytics (see, for example: Three Examples of Predictive Analytics in Clinical Practice; Diagnostic and Predictive Analytics and Their Possible Link to the Future of the LIS; Diagnostics Versus Therapeutics; The Changing Role of Physicians). On the basis of this new technology, pathology reporting to physicians will change radically in upcoming years. Pathology reports will no longer contain only a listing of test results or surgical pathology diagnoses but also a list of diseases and conditions that patients may develop in the future (see: Predisposition to Disease and Pre-Disease on the Health Continuum; The Challenge of Diagnosing Predisease in Our Healthcare Delivery System; The Diagnosis and Treatment of Pre-Diseases: A Look at the Future). It goes without saying that some patients will have no interest in their future diseases so such reports will certainly be modeled on clinician/patient preferences. 

This only rarely happens to me but the following passage from the article above caused me to laugh out loud: Chatbots will allow patients to ask questions about their EHR records. Although I can envision a small number of patients in the future wanting to query their EHR records with a chatbot, such a low level of interest  could not not possibly justify the development of such an app by Cerner and AWS. Much more likely as the impetus to develop such a chatbot would be to provide assistance to physician and nurses working with the EHR for extracting critical information from patient records.

But why, you might ask, would the article not cite such a justification for the development of "chatbot EHR query" as a feature highly desired by physicians and nurses? To state this overtly would constitute an admission that the EHR is difficult for medical professionals to use in its current state (see: The Benefits and Challenges of Electronic Health Records). Although not mentioned in the article above, I am also sure that an essential part of the development project turfed by AWS to Cerner is natural language processing (NLP) of the EHR record that will enable a larger portion of the record to be searchable using the aforementioned chatbots (see: AI Allows Computers to "Read" EHR Records and Make Predictions).

Diagnostics Versus Therapeutics; The Changing Role of Physicians

Prior to the 20th century, physicians themselves performed a small set of "lab tests" for their patients. One such example was tasting urine which was sweeter for diabetic patients (see: Urine flavor wheels helped doctors taste patients' pee centuries ago). Bedside diagnostic skills in these earlier days were more highly developed than today. Physicians of earlier generations thus performed all diagnostic and therapeutic functions for patients. This began to change in about the mid-20th century as the lab test menu expanded greatly and the sophistication of radiologic procedures increased. Commensurate with this shift, physicians' bedside diagnostic skills began to atrophy. Physicians caring for patients, of course, control all final diagnostic decisions and dictate all therapeutic interventions. However, what I will call the "lab/rad" contribution to final diagnoses is currently very large and will continue to increase.

Regarding the lab/rad contribution to final diagnoses and considering clinical pathology (CP), diagnoses are generally not rendered but rather key test results are reported that lead to final diagnoses. For anatomic pathology (AP), pathologists have always rendered precise diagnoses. Radiologists report "impressions" which continue to be more precise and, in many cases, actually constitute diagnoses. We are now on the verge of the rendering of definitive diagnoses in CP with increasing sophistication in molecular and genomic diagnostics. This trend will continue with increased diagnostic acumen of pathologists and radiologists driven by AI and deep learning. Add to this mix the increasing sophistication of wearable health ecosystems (see: The Evolution of "Wearable Health Ecosystems" and Associated Partnerships) such that healthcare consumers themselves now can generate increasingly sophisticated health data from home.

The conclusion from all of this is that physician-clinicians will continue to bear responsibility for making the final diagnoses for their patients after receiving and integrating critical reports form multiple diagnostic sources including lab/rad and, soon, connected home health (see: Connected health). This should not be construed as a problem but rather as an advancement in healthcare driven by science and technology. The downside of this trend is the increasing fragmentation and silo-ization of healthcare (see: The Antidote to Fragmented Health Care). Information technology can enable system interoperability and the merging of health records but this is a non-trivial challenge. 

Another Candidate for the Micro-Sample Diagnostics Market; CBC Results in Minutes

The Theranos fiasco (see: The rise and fall of Theranos, the blood-testing startup that went from Silicon Valley darling to facing fraud charges) served to cast a pall over the micro-sample diagnostics market. Now comes news about an Israeli company with a similar strategy but this time with an instrument providing a complete blood count (CBC) and a green light from the FDA (see: Unlike Theranos, startup’s blood test device ‘delivers on promise’ with FDA nod). Below is an excerpt from the article:

Israeli startup Sight Diagnostics has received US Food and Drug Administration clearance to market a blood test device that can extract the results of a standard complete blood count test from just a drop of blood in minutes....Sight’s OLO is a new type of blood analyzer that promises the same accuracy as large laboratory machines but is cost effective and can be used at the point of care, giving physicians the results “in minutes” ....To use Sight’s product, the physician or nurse pricks the patient’s finger and places a drop of blood into a disposable plastic cartridge that is inserted into the OLO, which looks like a small home printer. The machine, equipped with a camera, takes thousands of images of the millions of cells within the sample. Software developed by the firm based on machine learning algorithms analyzes the images and provides the results in a printout or via email....Recently, Sight also initiated a pharmacy pilot program with major UK chain Superdrug to bring blood testing to its health clinics, the company said in its statement. 

Basically, the Sight technology involves obtaining a couple of drops of blood from a patients and transferring them to a plastic cartridge that creates a "live" cellular monolayer.  This monolayer is then scanned and analyzed with machine vision and AI techniques to count and distinguish between the various types of cells in whole blood. This technology was initially developed by the company to diagnose malaria. Although the device is currently designed to provide only a CBC, it seems to me that the technology could be extended to diagnose various types of leukemias and other blood dyscrasias in the future.

One of the key points to consider here is where this new technology will be optimally deployed. For me, the options are: (1) central hospital labs; (2) point-of-care (POCT) testing locations such as hospital clinics or physician offices; or perhaps (3) retail drug store clinics. I would rule out central labs which already have sophisticated and highly automated hematology analyzers. The deployment of POCT devices in walk-in retail drug store clinics is still unusual in the U.S. I also think that the first such instruments to be deployed in these settings will be i-STAT type handheld devices for performing routine serum chemistry panels. I thus believe that the logical and preferred location for these OLO devices, if they prove to be efficient, accurate, and cost-effective, will be physician offices and hospital-based hematology, oncology, and surgery clinics. For them, rapid CBC results from such a devices on a POCT basis would be invaluable. Ultimately, I think that this device may also yield information about white cell morphology which will make them even more useful.

Plasma Proteins Used to Monitor Health Status; Emerging Science for Lab Testing

A recent article proposed a new way of analyzing patient serum as a means to broadly assess health status (see: Plasma protein patterns as comprehensive indicators of health). Below is an excerpt from the abstract:

Proteins are effector molecules that mediate the functions of genes and modulate comorbidities, behaviors and drug treatments. They represent an enormous potential resource for personalized, systemic and data-driven diagnosis, prevention, monitoring and treatment. However, the concept of using plasma proteins for individualized health assessment across many health conditions simultaneously has not been tested. Here, we show that plasma protein expression patterns strongly encode for multiple different health states, future disease risks and lifestyle behaviors. We developed and validated protein-phenotype models for 11 different health indicators: liver fat, kidney filtration, percentage body fat, visceral fat mass, lean body mass, cardiopulmonary fitness, physical activity, alcohol consumption, cigarette smoking, diabetes risk and primary cardiovascular event risk. The analyses were prospectively planned, documented and executed at scale on archived samples and clinical data, with a total of ~85 million protein measurements in 16,894 participants. Our proof-of-concept study demonstrates that protein expression patterns reliably encode for many different health issues, and that large-scale protein scanning coupled with machine learning is viable for the development and future simultaneous delivery of multiple measures of health. 

To state that the approach described in the paper is a totally new way to monitor health status is an understatement. In the study, about 5,000 different serum proteins were measured from about the 17,000 subjects, accounting for about 85 million assays. The authors labeled this new lab testing approach as a liquid health check. The protein-phenotype models were then developed using AI techniques. My eye was immediately caught by one of the possible uses of this liquid health check -- to identify excess alcohol consumption in a patient. Here an excerpt from an article on the common alcohol use biomarkers from ARUP (see: Alcohol Use Biomarkers): 

Approximately 20% of primary care patients in the United States drink alcohol (ethanol) at levels harmful to health....Diagnostic criteria for alcohol use disorder (AUD) vary, but the most widely used criteria are found in the fifth edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM-5). Laboratory tests for acute alcohol ingestion include ethanol, ethyl glucuronide (EtG), and ethyl sulfate (EtS) tests. Carbohydrate-deficient transferrin (CDT) and phosphatidylethanol (PEth) are useful markers for monitoring abstinence following long-term use.

Given that about one primary care patients in five has a serious drinking problem, it seems appropriate (or even necessary) to provide some sort of routine screening test and the liquid health check may provide such a test. Here's a quote from an article about clinicians' acumen in diagnosing this problem (see: Clinician Suspicion of an Alcohol Problem: An Observational Study From the AAFP National Research Network):

Clinician suspicion of alcohol problems had poor sensitivity but high specificity for identifying patients who had a positive screening test for alcohol problems. These data support the routine use of a screening tool to supplement clinicians’ suspicions, which already provide reasonable positive predictive value. 

In medical diagnosis, test sensitivity is the ability of a test to correctly identify those with the disease (true positive rate) whereas test specificity is the ability of the test to correctly identify those without the disease (true negative rate). Clinicians thus have trouble identifying patients with an "alcohol problem" based on a positive screening test but little trouble identifying non-alcoholics with a positive test.