BIG DATA: PHYSICIAN FRIEND OR FOE? Annual Health Law Conference - - PowerPoint PPT Presentation

BIG DATA: PHYSICIAN FRIEND OR FOE?

Annual Health Law Conference Northeastern University School of Law April 11, 2019

Barry Furrow Professor of Law and Director, the Health Law Program Kline School of Law @ Drexel University

• c. 2019

DATA ANALYTICS:

PROGRESSIVE HEALTH CARE TRANSFORMATION

• A. DATA USE: A HISTORY OF GENIUS, DELAYS, AND

PROGRESS

• B. DATA SOURCES: MORE DATA AND MORE

COMPUTERING POWER

• C. DATA ANALYTICS: PANNING FOR GOLD
• D. BENEFITS OF BIG DATA I: PHYSICIANS
• E. BENEFITS OF BIG DATA II: HOSPITALS
• F. BIG DATA HURDLES: A LAWYER/ CURMUDGEON

WORRIES

• G. SHOULD AI TAKE OVER? UNDER WHAT

REGULATORY CONSTRAINTS?

This supercomputer will perform 1,000,000,000,000,000,000 operations per second.

• A. DATA USE: A QUICK HISTORY OF

GENIUS, DELAYS, AND PROGRESS

Nightingale was a talented and creative statistician. She returned from the Crimea with extensive data on soldier mortality rates. She completed her 850-page book Notes on Matters Affecting Health, Efficiency, and Hospital Administration of the British. Her statistical analyses reformed health and data collection in both military and civilian hospitals. Nightingale transformed data visualization. She developed the graphic method -- the polar area graph -- to convey information about causes of death during the Crimean War. Each of the twelve wedges was then divided into three colors: blue representing deaths from contagious diseases such as cholera and typhus, red representing deaths from wounds, and black representing deaths from all other causes. At a glance, the vast majority of deaths were from contagious diseases, which were largely preventable.

• 1. Florence Nightingale – Master of Big Data,

Epidemiologist, Statistician, Graph Genius of Contagious Disease 1858

• 2. Ernest Codman, MD. Obsessive Collector of

Patient Data 1917

“You hospital superintendents are too easy. You work hard and faithfully reducing your expenses here and there-a half—cent per pound on potatoes or floor polish. And you let the members of the [medical] staff throw away money by producing waste products in the form of unnecessary deaths, ill-judged operations and careless diagnoses, not to mention pseudo-scientific professional advertisements.“

• 1917.

Dr. Ernest Codman-- ”End result survey”: every hospital should follow every patient that it treats, long enough to determine whether

• r

not the treatment has been successful, and then to inquire 'if not, why not?' with a view to preventing similar failures in the future.”

William Mallon, Ernest Amory Codman: The End Result of a Life in Medicine (Philadelphia: WB Saunders, 2000).

• Goal: A complete patient record to evaluate,

compare and establish benchmarks for the performance of physicians and hospitals.

CONCLUSION FROM THE SHORT HISTORY OF DATA IN HEALTH CARE:

HOSPITALS AND OTHER HEALTH CARE INSTITUTIONS ARE S L O W LEARNERS

• B. DATA SOURCES:

MORE DATA AND MORE COMPUTING POWER IS NOW AVAILABLE

AHRQ (Agency for Healthcare Research and Quality: Bringing Predictive Analytics to Healthcare Challenge

Learn about the $225,000 challenge to develop predictive analytics to estimate hospital inpatient utilization. Slow to the game and underfunded.

• C. DATA ANALYTICS: PANNING FOR

GOLD

KNOWLEDGE DISCOVERY PROCESS

DATA MINING—CORE OF KNOWLEDGE DISCOVERY PROCESS

Data Cleaning Data Integration Databases Data Warehouse Task-relevant Data Selection Data Mining Pattern Evaluation

DATA MINING LINKS THREE SCIENTIFIC DISCIPLINES

(1) Statistics: the study of data relationships using numbers; (2)

Artificial intelligence:

the use

• f

software and/or machines that display human-like traits; and (3) Machine learning: algorithms learning from data to make predictions.

• D. BENEFITS OF BIG DATA I:

PHYSICIANS

AI = PARTNER AND CONSULTANT.

AI HAS ALREADY OUTPERFORMED DOCTORS

• 1. Skin cancer diagnostics

58 dermatologists competed with a convolutional neural network. Doctors 86.6%; AI 95%. System developers used more than 100.000 pictures of this disease. Promise: early diagnosis of skin cancer.

• 2. Cancer and heart failure diagnostics

Lab developed a machine vision system for diagnosing different types of cancer (breast, prostate, head and neck), epilepsy and heart failure. System correctly predicted heart failure 97%; doctors 74%.

• 3. Pneumonia diagnostics

Stanford deep machine learning algorithm for chest X-ray images and detection of

• pneumonia. Can diagnose up to 14 types of medical conditions, and beat out 4 experienced

radiologists.

• 4. Early diagnostics of cardiovascular diseases

Oxford researchers system analyzes heart scans and predicts some diseases and possibility

• f heart attack more accurately than doctors.

• 1. Machine

learning will dramatically improve the ability of health professionals to establish a prognosis

• 2. Machine learning will improve diagnostic accuracy

Stanford algorithm, CheXNeXt, is the first to simultaneously evaluate X-rays for a multitude of possible maladies and return results that are consistent with the readings of radiologists, the study says. Scientists at Stanford have trained the algorithm to detect 14 different pathologies: for 10 diseases, the algorithm performed just as well as radiologists; for three, it underperformed compared with radiologists; and for one, the algorithm outdid the experts.

AI = CARE EXTENDER.

Telemedicine is obvious

• candidate. Here is a powerful and valuable use, coupling avatars

for therapeutic and diagnostics, reading films,

• ffering

suggestions for patient self-care and directing patients toward live doctors for live care.

VIRTUAL HUMANS and AVATARS

Virtual humans (VHs) improve clinical interviews, improving such screenings by increasing willingness to disclose information. Automated VHs can help overcome a significant barrier to

• btaining truthful patient information.

Research Report. Gale M. Lucas et al, It’s only a computer: Virtual humans increase willingness to disclose, Computers in Human Behavior 37 (2014) 94–100

• E. BENEFITS OF BIG DATA II:

HOSPITALS

AI = HUNTER FOR ERROR. Search for medical

adverse events in hospital, variations in practice below the mean level of infections, mortality, patient satisfaction. Report cards are then step 1; staff privilege actions are step 2. Must such data profiling be send to the National Practitioner Data Bank? If not, what should be sent as data mining reveals the poor doctors, the incompetents, the substandard?

DATA ANALYTICS WILL DISCOVER PHYSICAN- CAUSED ADVERSE EVENTS.

WILL HOSPITALS STEP UP THE HUNT FOR BAD “DOCS”? SHOULDN’T THEY?

AI Reduces Adverse Events

Adverse events among inpatients at three leading hospitals are 33.2 percent of hospital admissions for adults, up to ten times previous studies.

David C. Classen et al, ‘Global Trigger Tool’ Shows That Adverse Events In Hospitals May Be Ten Times Greater Than Previously Measured, 30 Health Affairs 581 (2011) (uses global trigger tool, a form of chart review that searches for triggers that mark adverse events.)

“Operation of a common automated ADE surveillance system across hospitals permits meaningful comparison of ADE rates in different inpatient settings. Automated surveillance detects ADEs at rates far higher than voluntary reporting…” Peter M. Kilbridge, Udobi C. Campbell, Heidi B.

Cozart, Maryam G. Mojarrad, Automated Surveillance for Adverse Drug Events at a Community Hospital and an Academic Medical Center

AI Identifies Hospital Acquired Conditions (HAC)

“Leveraging machine-learning, analytics can identify emerging complications and alert clinicians to prevent serious harm to patients, excessive costs and long-term healthcare needs.”

The focus on data has made a dramatic impact in contributing to a reduction in hospital acquired infections, which not only results in better delivery of care but also improved financial performance. Major causes for HACs have steadily declined since 2008.” Health Care Analytics: How Data is Changing Everything, Conduent Business Services 2017,

https://downloads.conduent.com/content/usa/en/ebook/healthcare- analytics.pdf

WILL DOCTORS AS EMPLOYEES BECOME MORE VULNERABLE? AS INDEPENDENT MEDICAL STAFF? WILL DATA ANALYTICS FINDINGS CIRCUMVENT PEER IMMUNITY STATUTES?

See Barry R. Furrow, Searching for Adverse Events: Big Data and Beyond, 27 Annals of Health Law 149 (2018)

AI = GATEKEEPER OF FUTILITY JUDGMENTS.

First comes determination of threshold for palliative care, reducing

• therwise standard treatments.

Sounds good. Then comes insurance incentives to find this out, to use it to demand less care,

• r even drop patients. Not likely?

80 percent of Americans would prefer to spend their last days at home if possible. But up to 60 percent die in an acute care hospital while receiving aggressive medical treatments. See Stanford paper “Improving Palliative Care with Deep Learning” published on the arXiv preprint server. Stanford’s algorithm is a Deep Neural Network trained on the EHR data from previous years, to predict all-cause 3-12 month mortality

• f patients as a proxy for patients that could benefit from palliative
• care. We want to predict when to stop treating. Recognizing futility

is important for the patient and the health care system. BUT….Imagine a health insurer using a death-forecasting

algorithm to make decisions about whether to pay for medical care. Is imminent death a preexisting condition?

DEATH-PREDICTING ALGORITHM

AI = STANDARD OF CARE SETTER.

AI creates Baselines, Benchmarks, and Awareness

• f

Provider Failures—So.... Retrain, Revoke, Remove

Hospitals can use data analytics to generate physician report cards, computer reminders to prescribe evidence-based medicines, or even clinical-decision support tools to help them assess patient risks for blood clots. Once hospitals compile data, they will be forced to develop new strategies of adverse event reduction, from support tools to ways of penalized physicians who continue to be poor performers.

The standard of care becomes more demanding as AI efforts improve.

Physician “report cards” for diabetes, one of the highest-prevalence conditions in medical practice, were unable to detect reliably true practice differences within the 3 sites studied. Use of individual physician profiles may foster a care environment in which physicians can most easily avoid being penalized by avoiding or deselecting patients with high prior cost, poor adherence, or response to treatments.

Timothy P. Hofer et al., the Unreliability of Individual Physician “Report Cards” for Assessing the Costs and Quality of Care of a Chronic Disease, 281 JAMA 2098 (1999)

WILL SUBSTANDARD PHYSICIAN FINDINGS BE REPORTABLE TO THE NATIONAL PRACTITIONER DATA BANK?

AI = PROVIDER SUBSTITUTE. Why not, beyond the

normal telemedicine uses? For the third world, rural areas, much

• f the world, it is much better than nothing. For the physician in

practice, less drudge work.

AI = BRAINS OF THE FULLY AUTOMATED HOSPITAL.

Can AI be hacked? Of course.

See Samuel G. Finlayson, John D. Bowers, Joichi Ito, Jonathan

• L. Zittrain, Andrew L. Beam, and Isaac S. Kohane, Adversarial

Attacks on Medical Machine Learning, 363 Science 1287 (2019)

IF AI MAKES AN ERROR THAT HURTS A PATIENT, WHO IS RESPONSIBLE? WHERE SHOULD WE LOCATE RESPONSIBILITY AND THEREFORE LIABILITY?

• 1. DOCTOR WHO RELIES ON IT
• 2. AI DESIGNER
• 3. HOSPITAL OR OTHER HEALTH CARE FACILITY

• F. BIG DATA HURDLES: MORE WORRIES

1.

• CAPTURE. Poor data habits: EHR data does not match patient

records; Value, governance of data needed 2. CLEANING. Dirty data must be scrubbed to be accurate, correct, consistent, relevant and uncorrupted. 3.

• STORAGE. Cost, security, performance. Cloud storage is common, but

hybrid with onsite is become more common

4.

SECURITY. Data security problems from cloud hacking, personnel habits 5.

• STEWARDSHIP. Long shelf life needed for storage of records.

6.

• QUERYING. For analytics, how to query is a key issue.

7. REPORTING. 8.

• VISUALIZATION. Good Data Presentations

9. UPDATING 10. SHARING

Jennifer Biesnick, Top Challenges of Big Data Analytics in Health Care

–Volume and complexity –Physician’s interpretation –Poor mathematical categorization –Canonical Form –Solution: Standard vocabularies, interfaces between different sources of data integrations, design of electronic patient records

1.

HETEROGENEITY OF MEDICAL DATA

• 2. PROBLEM OF SLOW HOSPITAL

ADOPTION.

• 1. Lack of interoperability
• 2. Interstate variation in privacy rules overlaid on top
• f HIPAA
• 3. Cost of custom datasets for big systems; had to

share data; data analytics officers are expensive and scarce. 4. Skepticism about hyperpromises of data analytics

• 3. POOR EXECUTION

4. COSTS OF EFFECTIVE DATA ANALYTICS REQUIRES A DATA ANALYTICS

• OFFICER. ARE THEY AVAILABLE AND AFFORDABLE?

• 5. CLOUD COMPUTING: SECURITY,

SECURITY, SECURITY!

RISKS OF DATA USE BY SOURCES

IBM Watson™ Utilization

WellPoint used IBM Watson to provide approval suggestions to nursing staff based on clinical and patient data. Watson trained with 18,000 historical cases, used hypothesis generation and evidence-based learning to generate confidence

• scored recommendations to help nurses make decisions about

utilization management. But when used at cancer treatment centers, scanning the medical literature and patient records to advise doctors on treatment plans. Three years later, Watson Oncology hadn’t lived up to expectations.

• 6. VENDORS OVERPROMISE AND

UNDERDELIVER

• G. SHOULD AI TAKE OVER? UNDER WHAT

REGULATORY CONSTRAINTS?

AI IS OFTEN RACIST, SEXIST, AND ELITIST

AI is 100% dependent on datasets, and health care datasets perfectly record a history of unjustified and unjust disparities in access, treatments, and outcomes across the U.S. Non whites experience worse outcomes for infant mortality,

• besity, heart disease, cancer, stroke, HIV/AIDS, and overall

mortality. Alaskan Natives suffer 60 % higher mortality than

• whites. AIDS mortality for African Americans is increasing. And

among whites, ther are hug geographic differences in outcomes an mortality.

The raw data to train AI models can perpetuate these disparities. They do not verify the accuracy of the underlying data given.

Will Knight, “The Dark Secret at the Heart of AI: No one really knows how the most advanced algorithms do what they do. That could be a problem,” MIT TECHNOLOGY REVIEW (April 11, 2017)

Posted April 2, 2019 https://www.regulations. gov/document?D=FDA- 2019-N-1185-0001