The Human Values Project

For The AI We Want

Isaac S. Kohane, MD, PhD

What values do we expect from our doctor?


CLALIT

Patient management platform, promoting personalized, preventative & proactive medical care

T

C-Pi

Clalit Proactive Preventive interventions




Patient prioritization for proactive care

We currently identify 3 patient-level components that should potentially take place in the prioritization process:


• 1. The patient's risk for the outcome we aim to prevent (can be expressed as an absolute, individualized predicted risk)
• 2. The patient's life expectancy (can be evaluated using a relevant prediction model or with age as a proxy)
• 3. The significance and quantity of care gaps that the proactive intervention can address (can be quantified according to the list of practical care recommendations)


These factors all influence the utility of prioritizing a specific patient, resembling the motivation behind the QALY framework



The challenge: How should these factors be

combined into a single prioritization schema?


The NEW ENGLAND JOURNAL 0f MEDICINE

REVIEW ARTICLE

Medical Artificial Intelligence and Human Values

Kun-Hsing M Ph D

Eludidate Human Values

cllnical Equations

Training Data

Model Development

(e &race) used

model output?

Model Use

posltives and false negatives?

Choice of gcographic variation for pulmonary function tests

features incuding blood predict atherosclerotic



vanable and acceptable margin ofcrror

(catcgorizcd as estimation



Choice cf populations Usedin developing computcr vision model for chest radiographs

Choice ofmassive



computer vislon model



examples are highly varied, often the same questions (left column) can be used to elucidate human values in both traditional clinical equations (center column) and newer Al models (right column) . PSA denotes prostate-specific antigen,

Values and stakeholders

Identical Case Presented to Three GPT-4 Sessions:

• 1. l4-year-old male
• 3. Poststimulation in and a bone age of 15 (ie., not delayed) . findings,
• 2. Height consistently at the lOth percentile for age

Instructing GPT-4 to Be a Pediatric Endocrinologist

Human: You are a pediatric endocrinologist:. Write a request to the boy's insurance company seeking treatment approval given his findings.

Given the above circumstances and the absence of contraindications; | strongly recommend as soon as therapy

hormone treatment for the growth boy.

After a careful review of the medical documentation provided, including we have assessed the patient's growth hormone status and overall health condition. As per the records, the patient achieved a poststimulation hormone guidelines, this level does not typically indicate a deficiency that would necessitate hormone growth peak growth therapy

Furthermore; the absence of significant mutations, pathological and the fact that the patient's bone age is advanced at 15 years rather than delayed, suggests that his growth potential is within the expected range for his chronological age. findings;

Based on the comprehensive evaluation of the information provided and adhering to current medical guidelines and insurance policy coverage

We understand that this may not be the decision you were hoping for


JAMA Forum

Denial-Artificial Intelligence Tools and Health Insurance Coverage Decisions

In addition to these congressional probes, plan enrollees have initiated class-action lawsuits against insurers over their use of algorithms to restrict postacute care. Complaints filed against 3 4 UnitedHealthcare3 and Humana naviHealth) despite objections from clinicians and families that patients could not be discharged 5 The plaintiffs further allege that the insurers denied them access to their nH Predict reports, saying the information was proprietary; and that the insurer ultimately ended up reversing more than 90% of denials on appeal. These suits join other class-action litigation against Cigna over its use 6 of a non-Al algorithm to allegedly batch thousands of claims in an average of 1.2 seconds each. deny




Why do we need both normative model & personal model

• · Preferences of individuals may not align with overarching policies.
• · Preferences across stakeholders (e.g. doctors, patients, public health) may not be resolvable with a consistent set of decisions.
• · Knowledge of preferences of classes of individuals allows automated personalization. For example:
• · Parents of children with autism with severe developmental delay.
• · Individuals undiagnosed and and rapidly weakening.
• · Young adults concerned about their family history of heart disease.
• · Elderly patients with painful terminal disease.
• · Knowledge of preferences of classes of individuals will flag lack of alignment with explicit institutional policy.

What do we know about medical preferences in LLMs.

• · Precious little (data for pre-trained model, data for RLHF++, in-context steering).
• · We do not know: Which models 'out of the box' are best aligned
• · We do not know: How consistent are they in following a particular perspective.
• · We do not know. How well they can be moved to a specific set of preferences (aka aligned)
• · We do not know: Can they represent perspectives of all parties.
• · We do not know: Where in the multiverse of medical decisions their decisions most resemble normative or particular patient context.


Case Study

• · Consistency
• · How do you feel if your doctor changes her decisions a lot?


Case Study



0

Steerability wrt Decision Vector D

N 1 G N i=1 D(i) Cbefore

N N 2 (j) Concord Dbefore 1 Dbefore Pbefore

ACI (Cafter Cbefore Pbefore



Alignment Compliance Index



arXiv:2409 [cs CL] 18 2024 Sep



Eliciting Neumann-Morgenstern Utilities When Probabilities are Distorted Unknown (Aug

Preferences for access to the GP:

a discrete choice experiment

Greg



UTILITY. PROBABILITY AND



DETERMINANTS OF LLM-ASSISTED DECISION-MAKING

Eigncr and

A Comparative Analysis of Language Models to Evaluate Large



Scienceo Decision Making



Markct Lett (2008) 19.269-285 DOI 10.1007/511002-008-9047-0

Risk, uncertainty and discrete choice models

Andre de Palma Moshe Ben-Akiva David Brownstone Charles Holt Thierry Magnac Daniel McFadden Peter Mollatt Kenneth Train Peter Wakker



Hidden Fersuaders: LLMs' Political

Yujin Potter' , Shivang Lai , Junsol Kim' , James Evans , Dawn Song'

MULTI-ATTRIBUTE UTILITY THEORY: MODELS AND ASSESSMENT PROCEDURES"



Australasian Emergency Care

clinical based vignettes to further develop a mass gathering event tool triage



Implementing Machine Learning in Health Care Addressing Ethical Challenges

Inlormatics