Artificial Intelligence

elshammahnora@[Link]
3FMHZKEXSB
in Healthcare
Fairness: Addressing the Ethical, Regulatory, and
Privacy Issues

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Learning Objectives
1. Learn how we are all responsible for minimizing inequities related to AI in medicine.
2. Learn the goals of responsible implementation of AI in medicine.
3. Learn how to explain to colleagues that prospective RCTs provide the facts about health impact and
are therefore the most effective approach to accomplish these goals.
4. Learn how excluding variables of protected classes from data collection, models, and sensitivity
elshammahnora@[Link]
3FMHZKEXSB
analysis can be counterproductive.
5. Learn how we can use AI to reduce biases.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Be Solution Oriented Regarding AI Fairness
● Many complex ethical issues surround the use of Artificial Intelligence (AI) in our society.
● In medicine, when AI models are improperly created, implemented, or evaluated, they are capable of
leading to false conclusions, flawed clinical decision support, and can perpetuate inequities.
● This is a large and active area of research and numerous scholarly articles and links are listed below for
further reading.
● The bottom line is that the key to progress in this area is using rigorous science to be evidence based so
elshammahnora@[Link]
3FMHZKEXSB

that AI tools improve health outcomes for all.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Don’t Break Into Jail
● Be intentional about evaluating algorithms for bias, inequity, and potential harm.
● Start with the assumption that the data, algorithms, and implementations of AI are problematic.
● Assume that the impact of AI will be unfair and privacy problems will arise.
● Then, collaborate with experts to use science and statistical analyses to discover ways to understand
the biases and minimize inequity.
● Devote time and funding to ensuring that your AI tool not only addresses the minimum standard of
elshammahnora@[Link]
3FMHZKEXSB

fairness but breaks new ground in the science of AI fairness.

● “One of the great things about AI is we should be able to use it both to identify disparities in healthcare
and also to raise the standards of care for everybody.” —Russ Altman, Stanford Professor of
Bioengineering.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Statistical Bias
● Bias can be introduced accidentally through faulty models or by unknowingly implementing systems
that systematically discriminate.
● Time and resources must be set aside to discuss the assumptions of the AI model and develop
appropriate solutions to potential problems.
● There are many forms of bias that can lead to misunderstandings in these discussions.
● In science and statistics, the technical definition of bias is a systematic error in the data collected
elshammahnora@[Link]
3FMHZKEXSB

(statistic) compared with the truth (parameter).

● For example, self-reported age differs from the true computed age (today’s date minus date of birth),
and it differs in a systematic way—an underestimate down to the whole number (always younger). This
differs from random errors, which could be underestimates or overestimates (younger or older).

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Racial Bias
● Complex racial and gender biases can be introduced and pose the most cause for concern.
● Unfair models implement predictors that systematically introduce bias, discriminate, and exclude.
● The recognition and replacement of unethical elements with appropriate predictors can improve AI
tools.
● Care should be taken to avoid indirectly amplifying and perpetuating biases and errors in the data.
● For example, including health literacy in a model could introduce a bias that prevents patients with low
elshammahnora@[Link]
3FMHZKEXSB

literacy from receiving appropriate prevention measures.

● When AI tools are not carefully designed and evaluated, they can use and propagate bias.
● Yet, advancements have been made in recent years in the science of AI fairness that have shown that
AI tools can be fairer and less biased than humans.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Legally Recognized “Protected Classes” or “Sensitive Attributes”
Based on U.S. Laws or Regulations
● Race/ethnicity
● Sex (including gender, sexual orientation, and gender identity)
● Age
● Physical or mental disability
● Religion
elshammahnora@[Link]
3FMHZKEXSB

● National origin or ancestry

● Citizenship
● Pregnancy
● Familial status
● Veteran status

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 The Evaluation of AI Fairness and Equity Must Be Ongoing During
the Entire Project and Is a Process That Is Best Managed With
Continuous Quality Improvement (CQI)
● Since AI fairness and equity issues are not an easy one-time fix, a CQI project provides the right
framework.
● The ethics of AI involves more than race and ethnicity.
● Ideally, each of the categories in the previous slide must be addressed.
elshammahnora@[Link]
3FMHZKEXSB

● Also, strive to assess for fairness and equity by insurance type, income level, education level, etc.
● Therefore, these variables must be in your data dictionary, recorded, and exported in your data set.
● The analyses can all be accomplished with a sensitivity analysis that you put in the appendix of your
paper. Willingness to adapt algorithms to improve fairness is essential.
● Although it is not feasible to address all these issues in any one project, this
can be viewed as a long-term goal, which is why the CQI approach is ideal.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Avoid “Colorblindness”—Removal of Race as a Predictor Is Not a
Solution
● Models should be trained on a diverse population.
● Publication of predictive models should include a detailed table showing the diversity of the population
and sensitivity analysis to assess how well the model performs overall and in diverse groups.
● Sensitivity analysis simply means that the results are evaluated and displayed for subgroups.
● This post hoc analysis provides information about the robustness of the results and conclusions.
elshammahnora@[Link]

● The Randomized Controlled Trial (RCT) should monitor how the model impacted outcomes in various
3FMHZKEXSB

groups.
● These issues must be studied, and the results transparently reported in papers. Of course, this requires
having a sample that is large and diverse enough to provide meaningful conclusions from these
subgroups.
● Not recording race and excluding race in all models is ignoring the problem,
not solving it.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Involve Stakeholders, Including Ethicists, Data Activists, Patients
and the Public in the Planning, Execution, and Reporting of the AI
Project
● Fairness and equity in AI are critical issues—not afterthoughts.
● The AI team must accept expert guidance from the communities that the system will affect and invest
time and resources to do this appropriately; AI can, and will, be used in ways that are ethical and fair
using deidentified data that are protecting personal information.
elshammahnora@[Link]
3FMHZKEXSB

● Design the study and the analysis with a primary goal of minimizing inequity in “protected subgroups”.
● Not recording (or exporting data on) these variables and “forbidding” these variables as predictors and
features in AI algorithms are not effective approaches.
● The solution is to ensure that measures of health are improved even as methods are automated at
scale—overall, and in these protected groups.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 The Elements of Fair and Responsible AI
● Ethical purpose: How will it benefit patients and society?
● Accountability: Who is responsible for the way it works?
● Transparency: How was the model developed, and how will it be implemented?
● Explainability: How does the algorithm work to convert inputs to outputs?
● Fairness and nondiscrimination: Is the model helping all groups?
● Safety and reliability: Is the model harmless and stable?
● Open and fair competition: Has the model been compared to alternative
elshammahnora@[Link]
3FMHZKEXSB

approaches? [Link]

●
ensuring-trust-in-modern-technology/
Privacy: How has the confidential patient data been protected?
● Robustness: Is the model empirically sound in different locations and across time?
● Beneficence: Is the model benefiting patients?
● Nonmaleficence: Has there been research to assess harm to or neglect of patients?
● Autonomy: Have patients been informed about decisions affecting their medical
care?
● Justice: Is there a fair and equitable distribution of burden and resources?
Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.
This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Concerns From the Public About Using AI in Medicine
● “I don’t want to have the computer making decisions instead of my doctor.”
● “My privacy may not be protected.”
● “The AI technology is not mature enough for medical applications.”
● “I don’t trust the AI companies.”
● “What about my choice as a patient?”
● “Will the model target or track me?”
elshammahnora@[Link]
3FMHZKEXSB

● “Will this increase my healthcare costs?”

● “The data are probably biased.”
● “Will my data be shared without my knowledge?”
● “This could widen the digital divide.”
● “AI could exacerbate existing racial disparities.”

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Be Transparent and Document Steps Taken to Improve AI
Fairness
● Variables, such as sex and gender, must be thoughtfully recorded and considered as predictors.
● Automatically omitting sex and gender from AI tools is not the solution.
● Men and women suffering from a heart attack report different symptoms in an emergency department.
● Ignoring gender in an AI tool to compute the probability of a heart attack would be foolish.
● AI should be optimized to use variables that are significant for all people. This needs to be part of the
elshammahnora@[Link]
3FMHZKEXSB

evaluation.
● Are there unintended consequences of including gender identity in a model? Do all subgroups show
improvement? Are there subgroups that are harmed by this? Are there age-gender interactions that
could be used as a predictor to improve the outcome for all groups?

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 How to Assess a Model for Bias
● Assess the area under the Receiver-Operating-Characteristic curve for each protected class or subgroup, and plot the Receiver-
Operating-Characteristic and calibration curves.
● Create a model for each protected class or subgroup and compare results.
● Assess the association between the protected class and the predictors in the model.
● Create a model to predict a protected class from the predictors and the endpoint in the model.
● Create a spline graph of the given prevention treatment on y vs. probability of the outcomes on x, with lines for protected classes.
● Create a spline graph of the outcome on y vs. probability of the outcomes on x, with lines for protected classes.
● Create a forest plot of the impact of the model on the protected classes.
elshammahnora@[Link]
3FMHZKEXSB

● Perform uncertainty quantification for subgroups and protected classes. Are there groups for which the 95% confidence interval
is just too wide?
● Create a classification and regression tree overall and for various protected classes to understand the model’s performance.
● Use explainable AI techniques to help solve bias issues.
● Finally, document and share with others your work in this area. Include a section in your paper or appendix showing your analysis
of this work to minimize inequities in protected classes.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 In Many Cases, It Is Unethical to Not Randomize and Not Learn
Which Approach Provides the Best Overall Health Results
● Attempting to stop AI projects until there are no biases in the data is unrealistic.
● Yet, there is a possibility that AI tools could worsen outcomes for patients or subgroups if scientific rigor is not used
in AI evaluations.
● Here, randomization can be justified from a safety perspective.
● Rigorous science will enable us to adopt AI in healthcare at the optimal pace—not too fast and not too slow.
● Do not overpromise the benefits of AI in medicine. False claims can be a deceptive act or practice that has legal
elshammahnora@[Link]
3FMHZKEXSB

ramifications.
● AI has amazing potential in medicine, but some of the declarations of victory have been premature. AI
entrepreneurs and enthusiasts are optimistic but claims that exaggerate the capability of AI risk falling into the
category of “deceptive practices” by the U.S. Federal Trade Commission (FTC). Racially biased algorithms may also
fall under this category. As the FTC states: “The FTC Act prohibits unfair or deceptive practices. That would include
the sale or use of, e.g., racially biased algorithms.”

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Summary
1. We are all responsible for minimizing inequities related to AI in medicine, and therefore, we must
search out potential causes of bias and address them.
2. The goals of responsible implementation of AI in medicine are to (1) improve health overall, (2)
improve health in protected classes, and (3) minimize inequality in outcomes.
3. Prospective RCTs provide the facts about health impact and are therefore the best approach to
elshammahnora@[Link]
3FMHZKEXSB
accomplish these goals.
4. Excluding variables of protected classes from data collection, models, and sensitivity analysis can be
counterproductive.
5. Responsible AI requires that clinical experts lead, remain in the decision loop, review the AI results,
and combine this with other information to make a fair and balanced decision.
6. If we do this right, AI will reduce biases.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 References
1. Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to
manage the health of populations. Science. 2019;366(6464):447-453.
2. Howard LM, Garguilo K, Gillon J, et al. The first 1000 symptomatic pediatric SARS-CoV-2 infections in
an integrated healthcare system: a prospective cohort study. BMC Pediatr. 2021;21(1):403.
3. Peters T. The Pursuit of Wow! Every Person’s Guide to Topsy-Turvy Times. Vintage: 1994.
4.
elshammahnora@[Link]
3FMHZKEXSB
Gabor A. The Man Who Discovered Quality: How W. Edwards Deming Brought the Quality Revolution
to America. Penguin Books; 1992.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Artificial Intelligence
elshammahnora@[Link]
3FMHZKEXSB
in Healthcare
Algorithm Bias in AI

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Algorithm (al-ge-ri-thəm) n. a procedure for solving a
mathematical problem in a finite number of steps that
frequently involves repetition of an operation.
elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 11 + 31 = 42
elshammahnora@[Link]
3FMHZKEXSB

What is ethical or responsible about this algorithm?

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Simpson’s Paradox
● Berkley gender bias in the 1970s.
● All departments admitted men at higher rates.
● The university admitted women at higher rates.
● Who is correct? Does gender bias exist?

elshammahnora@[Link]
3FMHZKEXSB
Women Men
Dept. A 0/1 < 50/100
Dept. B 70/100 < 1/1

Total 70/101 < 51/101

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Context Matters
elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Medical Outcomes
● Are white people healthier than black people?
● “Blacks have 26.3% more chronic illnesses than
Whites (4.8 vs. 3.8 conditions; P < 0.001).”
● Removing algorithm bias by substituting
“healthier whites” with “less healthy blacks”
until the “marginal patient is equally healthy.”
elshammahnora@[Link]
3FMHZKEXSB

● This creates substantial disparities in health

screening.
● “Blacks have more-severe hypertension,
diabetes, renal failure, and anemia, and higher
cholesterol.”

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Policing
● Predictive policing uses AI to
forecast crime likelihood and
proactively police areas.
● Data is typically drawn from prior-
arrest databases.
● This creates a feedback loop.
elshammahnora@[Link]
3FMHZKEXSB

● Potential bias in arrests.

Drug use est. by overdose data.

Does policing protect or target

the community it “serves”?

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Examining Bias

elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Examining Bias

elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Examining Bias

Successful loan earns $300

Unsuccessful loan loses $700
elshammahnora@[Link]
3FMHZKEXSB

Profit: -700 Profit: 20,300

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.
This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Bias Problems With “Fair” Strategies
● Max Profit (no constraint): The blue group is held to a higher standard than the orange group.

● Group-Unaware (same credit score threshold): The orange group will get fewer loans than the blue
group.

● Demographic Parity (same loan rate to each group): Fewer qualified people in the blue group are given
elshammahnora@[Link]
3FMHZKEXSB

loans.

● Equal Opportunity (same true positive rate): Don’t know who will and will not pay back a loan in
advance.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Sampling Bias
● Scenario: A tech company is developing an AI-based
facial recognition system for gender and uses a dataset
predominantly composed of images from public figures
and celebrities.
● Bias: This dataset is likely to underrepresent older
individuals, people of varying attractiveness, and ethnic
elshammahnora@[Link]
3FMHZKEXSB

minorities. As a result, the AI model trained on this

dataset may perform poorly when recognizing faces
outside these demographic groups.
● Implication: The facial recognition system may exhibit
significant inaccuracies and higher error rates for
underrepresented groups, leading to biased and
unreliable results in practical applications.
Buolamwini, J., & Gebru, T. (2018, January). Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on fairness, accountability and
transparency (pp. 77-91). PMLR.
Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.
This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Survivorship Bias
● Scenario: During WWII, returning aircraft were analyzed for
where to add armor. They observed damage on the wings and
fuselage, and thus suggested reinforcing these areas.
● Bias: This analysis only included planes that survived and
returned from missions. The missing data were from planes that
were shot down and did not return, which might have been hit in
elshammahnora@[Link]
3FMHZKEXSB

critical areas like the engines or cockpit.

● Implication: Focusing on the surviving aircraft led to incorrect
conclusions. The real vulnerabilities were in the parts that, when
hit, caused planes to be lost.

Mangel, M., & Samaniego, F. J. (1984). Abraham Wald's work on aircraft survivability. Journal of the American Statistical Association, 79(386), 259-267.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Bias
● Many types of bias
● 50 types of cognitive bias

elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Managing Bias Tradeoffs: A.D.A.P.T.
● Awareness: Know and measure different potential biases.

● Diversity: Contrast different metrics to observe whether they are providing consistent or divergent
information.

● Analysis: Holistically investigate potential reasons/causes for contrasting measures.

elshammahnora@[Link]
3FMHZKEXSB

● Publish: Publish your bias data and results of analysis to create transparency and explainability of
decisions. Remain open to criticism and alternate explanations.

● Transform: Implement changes based on feedback and continuous learning to

refine the model and mitigate identified biases, ensuring a dynamic and
responsive approach to managing bias.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Artificial Intelligence
elshammahnora@[Link]
3FMHZKEXSB
in Healthcare
EHR: Exporting, Cleaning, Managing Datasets,
and Integrating Models into the Electronic
Health Record

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Learning Objectives
1. Learn how to continuously improve the completeness and accuracy of the data in the EHR—the
Information Architecture.
2. Learn why more of the medical information needs to be stored in a usable, structured, and coded
format.
3. Learn why long-term, patient-centric measures are needed to create AI tools that matter.
4.
elshammahnora@[Link]
3FMHZKEXSB
Learn how leaders can enable teams to build and improve upon previous models.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 The EHR Systems Were Designed to Improve Hospital Billing Not
to Be Data Sources for Research or AI
● Electronic Health Record (EHR) information is plentiful, but healthcare leaders must prepare for the
future by investing in resources to build a strong information architecture to ensure data reliability,
completeness, and usability.
● “There is no AI without IA—Information Architecture.” —Seth Early
● Information Architecture (IA) refers to a system for storing, structuring, organizing, and labeling data in
elshammahnora@[Link]
3FMHZKEXSB

a format that can be easily used. Interoperability is the ability of computer systems or software to
exchange and make use of information.
● The percent of patients with data on body mass index, smoking status, and exercise level should be
tracked over time and improved as a continuous quality improvement project to gradually improve the
usefulness of the data.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Exporting Data From the EHR Is a Common Bottleneck in Building
AI Tools
● Most of the current EHRs record raw unstructured data but fail to combine the information to make it
more useful and predictive.
● This would be analogous to a weather person on TV who simply reports the current local temperature
and wind speed but does not combine any of the information into weather models.
● By combining large amounts of structured, coded weather data in models, the weather person can
elshammahnora@[Link]

provide accurate and precise forecast predictions.

3FMHZKEXSB

● This is the value that they provide and the next step for healthcare.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Data Is the New Oil, but Like Oil, It Needs to Be Processed Before
It Is Useful
● Data management is an important part of the process of transforming EHR data into a useful AI tool.
● Successful healthcare organizations recognize the need for this specialty and are investing in recruiting
and paying talented experts in this area.
● These informatics experts are skilled at merging files, recoding text into codes, and handling missing
data in a sophisticated way.
elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Successful Healthcare Organizations Will Quickly and Efficiently
Export Data, Build Models, and Add Them Into the EHR
● It is technically possible to:
1) Export a data set needed to create a predictive model from the EHR in 1 day.
2) Create an excellent predictive model in 1 day.
3) Code the model into the EHR in 1 day.
● Most models can be tested in a pragmatic Randomized Controlled Trial in about 1 year.
elshammahnora@[Link]
3FMHZKEXSB

● But this process can take forever. To fix this, forward-thinking healthcare leaders will decide that this is
a priority and put resources into it. They will hire skilled, experienced people to be on the AI
development team. These people will not try to accomplish this work in their spare time when they
have no other demands on them, but will be devoted to these projects. The leaders will be very hands-
on to move these projects along and help overcome obstacles.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 EHRs Should Collect More Structured Data
● Is this patient a current cigarette smoker? 0—no, 1—yes.
● If yes, number of packs per day.
● If yes, number of years smoking.
● If no, is this patient a former smoker? 0—no, 1—yes.
● If a former smoker, years since quitting.
elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 EHRs Should Evolve to Record More Information With Structured
Formatting

elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Capture Information in the Health Records That Is Linked to
Patient-Centered Short-Term and Long-Term Health Outcomes
● “If you cannot measure it, you cannot improve it.” —Lord Kelvin
● One of the major shortcomings of the current EHRs is a lack of short-
term and long-term outcomes data.
● Record outcomes that are important to the patient, not simply a
process metric or what is important to the physician or hospital.
elshammahnora@[Link]
3FMHZKEXSB

● For example, a hospital may consider patient care a success if they are
discharged alive, but the solution is to measure success from the
patient’s perspective, as Wes Ely, M.D., pointed out in his excellent
book, “Every Deep-Drawn Breath: A Critical Care Doctor on Healing,
Recovery, and Transforming Medicine in the ICU”.
● Much of the time, doctors have limited data about whether their
decisions led to improvements for their patients. This feedback loop is
essential.
Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.
This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Important Patient-Centric Outcome Variables That Should Be
Routinely Collected
● Survival: Survived/died, date of death, or date of last follow-up
● Hospital readmission: Often missing for readmissions to other hospitals
● Activities of daily living
● Quality of life
● Days to return to prehospital status: Work or school
● Functional status
elshammahnora@[Link]
3FMHZKEXSB

● Level of pain
● Patient satisfaction
● Quality-Adjusted Life Year (QALY): These range from 1 (perfect health) to 0 (dead); one year in perfect health is
equal to 1 QALY
● Patient Reported Outcome Measures (PROMs): Quality measures derived from patient-reported information
● Overall health score: (0 = dead to 100 = optimal health)

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Successful Biomedical Informatics Leaders Will Support Pragmatic
RCTs
● “We believe there is an urgent need to promote the use of RCTs of Clinical Information Systems, given
continuing reservations in the medical informatics community.”
—Joseph Liu and Jeremy Wyatt
● “The RCT is an important and powerful method, and an underused one in medical informatics.”
elshammahnora@[Link]
3FMHZKEXSB
—Joseph Liu

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Creating One General AI Predictive Model for a Health Condition
or Medical Complication Across the Entire Hospital Is Often
Possible and Preferable
● The assumption that each department or specialty would require a different model is often false.
● In our work building predictive models of hospital complications, one overall model performed as well as the
models designed for a subgroup of patients.
● For most conditions, it is not feasible to create a model for each department or unit in the hospital when one
elshammahnora@[Link]
3FMHZKEXSB

model could be used for the entire adult hospital.

● The work that would be required to code and maintain multiple models in the EHR for the same condition would
not be worth the effort.
● Separate models will, however, be needed for certain situations, such as the children’s hospital.
● If a group insists on a specific model for their own unit, they must demonstrate why this additional expense is
justified.
● In general, a variable can be added to a model, such as trauma, rather than creating
another model for trauma patients.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 For Implementation to Be Successful, Use the Continuous
Probability of Outcome and Avoid Dichotomizing Risk
● Display the continuous probability of an AI predictive model and do not simplify to a dichotomized alert
as high/low, or an ordinal alert—green/yellow/red.
● A surprising number of predictive models and scoring systems commit the double sin of dichotomizing
twice.
● They will dichotomize the input predictors (BMI = obese) and dichotomize the output probability (high
elshammahnora@[Link]
3FMHZKEXSB

risk for blood clot).

● This is an unnecessary waste of valuable information.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Convert Statistical Output Into a Usable Formula
To compute the probability of a Venous Thromboembolism (VTE) for a given
patient, one would multiply the patient’s values
by the coefficients. For example, age would be multiplied by 0.056.
The probability of a VTE would be 1/(1+〖2.7183〗^(−Z)) × 100.
Round to one decimal place and display as a percent. Where Z = −2.537 + (2.156 ×
1 if patient has a history of thrombosis, 0 otherwise) + (0.045 × 1 if surgical
procedure performed during this encounter, 0 otherwise) + (0.486 × 1 if patient
has a diagnosis of cancer, 0 otherwise) + (0.932 × 1 if infectious disease consult
elshammahnora@[Link]
3FMHZKEXSB

ordered this encounter, 0 otherwise) +

(1.436 × 1 if cardiology consult ordered this encounter, 0 otherwise) + (1.121 × 1 if
blood gas lab panel was ordered this
encounter, 0 otherwise) + (1.583 × 1 if patient has a central line present, 0
otherwise) + (0.056 × patient age in years) +
(−0.147 × patient mean corpuscular hemoglobin concentration (MCHC). If MCHC is
missing, impute a value of 34.0. +
(0.059 × patient red cell distribution width (RDW). If RDW is missing, impute a
value of 14.3) + (0.079 × patient lactate.
If lactate is missing, impute a value of 1.3).
Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.
This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Enable Users to Notify the Model Developer/Researchers of
Errors and Use This to Adapt and Improve the Model
● Adding a mechanism that would allow the user to send feedback, such as “This does not look right,” is
a way to build in a feedback loop and to continuously improve and correct faulty models.
● Feedback might be as vague as “The model says X, but I think it should be Y.”
● This can also be used to flag a potential bias that should be investigated.
● Additionally, develop a feedback loop for understanding when and why the clinician ignores the AI
elshammahnora@[Link]
3FMHZKEXSB

recommendation.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Use Augmented Vigilance With Automatic Escalation in a Closed
Loop System to Be More Effective Than the Average Alert
● If clinicians ignore the alert, there will be an appropriate escalation.
● The goal is to improve end-to-end performance.
● If the clinician changes a behavior based on the alert, there is no escalation, but if not, the next level up
alert is prompted.
● This process continues until the intervention is performed or until there is a valid reason documented
elshammahnora@[Link]
3FMHZKEXSB

for not doing so.

● For example, a car’s seat belt alert uses a form of escalation that moves from a red icon to a beeping
sound to an annoying and continuous alarm.
● Implementation research is extremely important and often given inadequate attention.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 After AI Tools Are Implemented, Have a Post-deployment
Surveillance to Assess for Performance Drift
● Periodically, assessment of the model’s performance should be made by a standing data and safety
monitoring board with an auditing of the inputs and outputs for a random sample of patients.
● Remember, you are chasing a moving target, and models need to be revalidated periodically to ensure
that they still perform as expected.
● Here, calibration drift detection systems can be useful for deciding when to update the model.
elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Ironically, the Goal Is to Break the AI Model
● The goal is to prevent a complication in the patient predicted to have one.
● In this way, the outcome will change in the intervention group compared with the control group.
● Ironically, the long-term goal of a predictive model project is to have the model become obsolete!
● The implementation team is constantly trying to prove that the model is wrong.
● If the model successfully predicts when patients will develop a condition or complication, and the
hospital resources are focused on preventing that high-risk patient from developing that complication,
elshammahnora@[Link]
3FMHZKEXSB

then prospectively, the model will appear to no longer predict as well as before or as expected.
● Of course, this will not happen in the control group, and therefore we can assess cause-and-effect
during the trial.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Build Models That Are Robust to Dataset Shifts; Brittle Models
Are a Major Problem in AI
● Brittleness is the tendency of a model to be easily fooled with slight changes in the data.
● This is the opposite of robustness and stability.
● Brittleness causes the model to fail to generalize in other settings for a variety of reasons: different
equipment, different practice patterns, different names for predictors, etc.
● Dataset shift refers to the fact that within a healthcare system and EHR, the model that was built on
elshammahnora@[Link]

data from years ago may not perform appropriately today.

3FMHZKEXSB

● The solution to this problem is to reassess the model periodically. Is the rate of the outcome similar
today as when the model was built? Are these predictors still the most important factors?
● Building robust and fair models requires evaluations to address accidental fitting of confounders,
unintended discriminatory bias, the challenges of generalization to new populations, and the
unintended negative consequences of new algorithms on health outcomes.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Vaporware of What a Primary Care Provider Will See in the EHR
at an Annual Checkup

elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Summary
1. As hospitals increasingly rely on EHRs for not only billing but also research and real-time AI tools, more
effort needs to be devoted to continuously improving the completeness and accuracy of the data in
the EHR—the IA.
2. Common data elements—standardized key terms—have been discussed for decades, but strong
leadership and funding are needed for more progress in this area.
3.
elshammahnora@[Link]
3FMHZKEXSB
More of the medical information needs to be stored in a usable, structured, and coded format.
4. Long-term, patient-centric measures are needed to create AI tools that matter.
5. Leadership must enable teams to build and improve upon previous models.
6. The wasted resources of creating hundreds of unused models must be focused on building models 102
and 103, to keep improving on existing overall models.
7. Oversight from healthcare leaders is needed to guide the portfolio of AI work,
building on previous successes and moving forward.
8. Without this, researchers will continue to create models that are never
implemented.
Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.
This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 References
1. Ely W. Every Deep-Drawn Breath: A Critical Care Doctor on Healing, Recovery, and Transforming
Medicine in the ICU. Scribner; 2021.
2. Liu JL, Wyatt JC. The case for Randomized Controlled Trials to assess the impact of clinical information
systems. J Am Med Inform Assoc. 2011;18(2):173-180.
3. Liu JL, Wyatt JC, Deeks JJ, et al. Systematic reviews of clinical decision tools for acute abdominal pain.
elshammahnora@[Link]
3FMHZKEXSB
Health Technol Assess. 2006;10(47):1-167, iii-iv.
4. Davis SE, Greevy RA, Jr, Lasko TA, Walsh CG, Matheny ME. Detection of calibration drift in clinical
prediction models to inform model updating. J Biomed Inform. 2020;112:103611.
5. Davis SE, Greevy RA, Fonnesbeck C, Lasko TA, Walsh CG, Matheny ME. A nonparametric updating
method to correct clinical prediction model drift. J Am Med Inform Assoc. 2019;26(12):1448-1457.
6. Kattan M. Cleveland Clinic risk calculator library. [Link]

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Artificial Intelligence
elshammahnora@[Link]
3FMHZKEXSB
in Healthcare
AI Regulation

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 What Concerns You the Most?
● Rank the following organizations from most to least concern regarding use of your data.
● Now rank the organizations based on use of AI for making decisions regarding you.

elshammahnora@[Link]
3FMHZKEXSB

Government Policing/Judiciary Banking/Finance Medicine Human Resources

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 The AI Arms Race

elshammahnora@[Link]
3FMHZKEXSB

“Whoever becomes the leader in China is developing autonomous lethal N. Korea’s suicide drones
[AI] will become the ruler of the drones at scale. Three warfare: kinetic,
world.” —Vladimir Putin, 2017 legal, psychological

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Two Implications of the AI Arms Race
1. AI advancement is inevitable:
○ Adversaries will not wait—ethical or not, they will deploy it.
○ Therefore, democratic societies should use AI to enhance healthcare, mental health, public safety,
and quality of life.

2.
elshammahnora@[Link]
3FMHZKEXSB
Regulatory asymmetry as a weapon:
○ Propaganda will aim to convince open societies to adopt over-regulation, stifling innovation and
giving adversaries a strategic edge.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Current State of AI Regulation
● poorly defined in law:
○ Existing laws are outdated, failing to reflect AI’s learning, adaptive nature

● Two dominant regulatory frames:

○ Data privacy and security (e.g., GDPR)
elshammahnora@[Link]
3FMHZKEXSB
○ Narrative control and content moderation (e.g., social media regulations)

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 What Does It Mean to “Control” AI?

Large Model
Accuracy: 98%
Runtime: 1.5 sec
Size: 200 MB

elshammahnora@[Link]
3FMHZKEXSB
Smaller Model
Accuracy: 96%
Runtime: 0.2 sec
Size: 20 MB

● Example: GPT-4 cost >$100M to train; distilled alternatives < $5M.

● Key Point: The value lies not in the app but in the model weights. Who owns
them? Who has access? Who can use them?

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Competing Views on Ownership/Control

elshammahnora@[Link]
3FMHZKEXSB

Corporate Control (Capitalism): Open Access (Anarchy-Democracy): Government Control (Socialism):

• Argument: Tech giants have • Argument: Open-sourcing allows • Argument: Centralized regulation ensures
resources and capabilities. public innovation and public interest/safety.
• Counterpoint: Corporations are accountability. • Counterpoint: Risk of political misuse and
unaccountable, driven by profit, • Counterpoint: Opens the door to suppression of dissent.
and may be corrupt. malicious use (e.g., deepfakes,
fraud, cyberweapons).

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 The Fundamental Dilemma
● “If you can’t trust corporations (capitalism), can’t trust people (anarchy), and can’t trust government
(socialism)—then who can you trust?”
● Most regulation debates are not binary—they lie along a spectrum of trust and control.
● The key is understanding where different societies draw the line.

China: Gov’t first regulation; companies/individuals monitored

EU: Strong GDPR privacy laws; cautious innovation stance
elshammahnora@[Link]
3FMHZKEXSB

Singapore/Japan: Central oversight, limited freedoms

U.S./India: Deregulation promoting small business innovation

Europe U.S. Notably Absent: A pure capitalist approach

India
China

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Unintended Consequences of Regulation

Facial Expression Recognition (FER)

elshammahnora@[Link]
3FMHZKEXSB
and eye tracking measures mirror neurons

Avatarization creates too

much information loss
Three-brain hypothesis

MetaQuest Pro

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 What Concerns You the Most?
● Rank the following organizations from most to least concern regarding use of your data.
● Now rank the organizations based on use of AI for making decisions regarding you.

elshammahnora@[Link]
3FMHZKEXSB

Government Policing/Judiciary Banking/Finance Medicine Human Resources

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 What should regulation aim for?
elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Artificial Intelligence
elshammahnora@[Link]
3FMHZKEXSB
in Healthcare
Case Study: Predicting Risk of Systemic
Autoimmune Disease
PI – Dr. April Barnado

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Learning Objectives
1. Learn how to create an AI tool to predict the risk of systemic autoimmune disease in patients with
positive Antinuclear Antibodies.
2. Learn how to obtain research funding for AI tools in healthcare.
3. Learn how to test a model in a pragmatic trial.

elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Funding for the ANA Risk Model
● Predicting risk of systemic autoimmune disease in patients with positive ANAs:
○ April Barnado, MD, MSCI (Rheumatology/Medicine)
○ Dan Byrne, MS (Biostatistics)
○ Ryan Moore, MS (Biostatistics)
○ Marc Beller (Health IT)
● Submitted R01 Katz Award to NIAMS, Impact score: 33, 11th percentile:
○ Current funding 16th percentile for early-stage investigators, 12th percentile for
elshammahnora@[Link]
3FMHZKEXSB

competing renewals
○ 5 years (starting 2/1/22)
○ ~$250,000/year (direct costs), ~$460,000/year (total costs)
● Selected for VUMC Department of Biomedical Informatics Catalyzing Informatics
Innovation (CI2) Program:
○ 9/27/2021 – 9/26/2023 (2 years), $20,000 each year
○ Provides mentorship and work-in-progress meetings for scholars
○ “Learn to use real-world data and evidence to advance innovation, evaluate
methods, and impact change” (Program leadership: Daniel Fabbri and Bill Stead)
Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.
This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Our Model Paper

elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Objective
● Positive Antinuclear Antibodies (ANAs) cause diagnostic dilemmas for clinicians.
● Currently, no tools exist to help clinicians interpret the significance of a positive ANA in individuals
without diagnosed autoimmune diseases.
● We developed and validated a risk model to predict the risk of developing autoimmune disease in
positive ANA individuals.
elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Methods
● Using a de-identified Electronic Health Record (EHR), we randomly chart reviewed 2,000 positive ANA
individuals to determine if a systemic autoimmune disease was diagnosed by a rheumatologist.
● A priori, we considered demographics, billing codes for autoimmune disease-related symptoms, and
laboratory values as variables for the risk model.
● We performed logistic regression and machine learning models using training and validation samples.
elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Results
● We assembled training (n = 1030) and validation (n = 449) sets.
● Positive ANA individuals who were younger, female, had a higher titer ANA, higher platelet count,
disease-specific autoantibodies, and more billing codes related to symptoms of autoimmune diseases
were all more likely to develop autoimmune diseases.
● The most important variables included:
○ Having a disease-specific autoantibody
elshammahnora@[Link]
3FMHZKEXSB

○ Number of billing codes for autoimmune disease-related symptoms

○ Platelet count
● In the logistic regression model, AUC was 0.83 (95% CI 0.79-0.86) in the training set and 0.75 (95% CI
0.68-0.81) in the validation set.

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Conclusion
● We developed and validated a risk model that predicts risk for developing systemic autoimmune
diseases and can be deployed easily within the EHR.
● The model can risk stratify positive ANA individuals to ensure high-risk individuals receive urgent
rheumatology referrals while reassuring low-risk individuals and reducing unnecessary referrals.

elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Timeline of Model Covariates

elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Flow Chart of Training and Validation Sets

elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Characteristics of Incident Positive ANA Individuals in Training
and Validation Sets

elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Characteristics of Positive ANA Individuals With vs. Without
Systemic Autoimmune Disease in the Training Set

elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Importance of Variables in ANA Risk Model

elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Probability of Systemic Autoimmune Disease Based on Sex and
Age

elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Model Performance for Training and Validation Sets

elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Shiny App
● [Link]

elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Formula for Systemic Autoimmune Disease Risk Model

elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Comparison of Importance of Variables in Logistic Regression and
Cox Models

elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Summary
1. We created an AI tool to predict the risk of systemic autoimmune disease in patients with positive
Antinuclear Antibodies.
2. The PI was successful in obtaining research funding for AI tools in healthcare.
3. The model is being tested in a pragmatic trial.

elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 Thanks To
● April Barnado MD, MSCI
● Ryan P. Moore MS
● Henry J. Domenico MS
● Sarah Green BA
● Alex Camai BA
● Ashley Suh BA
elshammahnora@[Link]
3FMHZKEXSB

● Bryan Han
● Katherine Walker
● Audrey Anderson
● Lannawill Caruth BA
● Anish Katta BA
● Allison B. McCoy PhD

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.
 References
● Barnado A, Moore RP, Domenico HJ, Green S, Camai A, Suh A, Han B, Walker K, Anderson A, Caruth L,
Katta A, McCoy AB, Byrne DW. Identifying antinuclear antibody positive individuals at risk for
developing systemic autoimmune disease: development and validation of a real-time risk model. Front
Immunol. 2024 Mar 20;15:1384229. doi: 10.3389/fimmu.2024.1384229. PMID: 38571954; PMCID:
PMC10987951.
elshammahnora@[Link]
3FMHZKEXSB

Proprietary content. ©All Rights Reserved. Unauthorized use or distribution prohibited.

This file is meant for personal use by elshammahnora@[Link] only.
Sharing or publishing the contents in part or full is liable for legal action.