Ayush Noori and Reza Shamji

Zitnik Lab, Harvard Medical School

The Future of Personal AI Workshop

@ayushnoori @reza_shamji

An adaptive clinical �copilot for global health

We introduce ARK, a knowledge-grounded clinical AI model that adapts in response to expert feedback and context.

ARK�AI model

Memory module

“The combination of high blood pressure (145/95), protein in urine (+2), and neurological symptoms (headache, blurry vision) suggests Amani’s condition could rapidly progress to eclampsia.”

Clinician review

Clinical accuracy: 5/10

Possibility of harm: 1/5

Feedback: “Pre-eclampsia can often be confused with severe malaria, which can also cause headaches.”

Knowledge aggregator

Rationale

1

2

3

4

5

The memory module is a traceable KG with metadata; every fact is linked back to a clinical guideline or expert feedback.

Initialize with clinical guidelines

Update based on clinician feedback

Memory module

Node: No access to pulse oximeter�Source: Clinician report #49

62

Node: Severe childhood pneumonia

Source: WHO IMCI and IMNCI guidelines

762

Node: ORS non-response, malnourished�Source: Clinician report #82

1509

Node: Oral rehydration therapy (ORS)�Source: WHO diarrheal disease guidelines

1614

Node: Antibiotic use for febrile illness�Source: Clinician report #91

42893

Node: Hypertension diagnostic threshold�Source: OpenEvidence

7141

Humans can assess the output quality of ARK across diverse axes of evaluation.

Helpful rationale

�Is the model’s rationale helpful in determining whether the answer is correct?

Possibility of harm

​

Based on model’s output, is there a risk that the recommendation could cause clinical harm?

Clinical consensus

�Does the answer reflect established clinical practice and standard-of-care medical guidelines?

Completeness

�Does the model provide a complete response covering all necessary elements?

Task success

�Did the model successfully complete the diagnostic or therapeutic task it was given?

Cognitive traceability�

Are intermediate reasoning steps and decision factors interpretable and traceable?

Accuracy

�Are there any factual inaccuracies or irrelevant information in the response?

Clinical relevance

​

Does the model focus on clinically meaningful aspects (patient groups, relevant outcomes)?

Based on expert feedback, ARK can be periodically post-trained and redeployed.

ARK�AI model

RLHF

Redeployment in LMIC clinic