The Missing Chart

Inspiration

As a doctor, I understand that clinical reasoning rarely begins with a complete picture. Patient information arrives over time, and clinicians constantly reconsider their thinking as new symptoms, medications, laboratory results, and outside records become available.

Through my work across healthcare and AI, I have also seen a persistent divide. Many clinicians hesitate to trust AI because it can produce confident responses based on incomplete context or misunderstand clinical nuance. At the same time, AI engineers may understand the technology without fully understanding how medicine is practiced or how clinical decisions evolve.

I built The Missing Chart to help bridge that gap. It reflects how I think as a clinician: reason with what is currently known, recognize what is missing, verify consequential claims, and revise the assessment when the evidence changes.

What it does

The Missing Chart is an interactive clinical education simulation built around an evolving synthetic patient record.

Learners review the information available at a specific point in care, decide what to include in a focused request, examine the resulting assessment, and identify claims that require verification. As new records become available, they reassess the patient and see how missing context can change the interpretation, urgency, and appropriate next steps.

Rather than treating AI output as a final answer, the simulation teaches learners to examine the relationship between the information supplied and the response received.

How we built it

I designed the clinical scenario, educational objectives, evolving-information structure, and verification activities using my background in medicine, health informatics, and healthcare AI evaluation.

I used Codex to translate that clinical concept into a working application, including the interface, timeline progression, record-selection logic, deterministic response pathways, accessibility improvements, testing, and repository refinement. GPT-5.6 supported clinical-language refinement, safety review, debugging, and evaluation of the learner workflow.

The application was built with Next.js, React, TypeScript, and Node.js and deployed through OpenAI Sites. Its deterministic structure gives every learner a consistent experience without requiring an API key or credits.

Challenges we ran into

The central challenge was representing how clinical information evolves without making the experience confusing. The application needed to distinguish between information that exists somewhere in the patient’s record and information that is actually available to the clinician at that moment.

Another challenge was making the generated assessment useful without allowing it to appear complete or authoritative. The experience had to show that an output may be internally reasonable and still be clinically insufficient because important context was never included.

I also had to translate clinical reasoning into explicit application logic. Clinicians often integrate uncertainty, missing evidence, and changing priorities intuitively. Turning that process into a structured and understandable simulation required careful design.

Accomplishments that we're proud of

I am proud that The Missing Chart represents a recognizable clinical problem rather than a simplified question-and-answer exercise.

Information selection directly affects the assessment shown to the learner. The simulation makes missing context visible and allows learners to compare an initially reasonable response with a revised assessment after new evidence appears.

The application also brings clinical education, AI literacy, safety, and evaluation into one workflow. It uses entirely synthetic patient information, requires no login or credentials, and keeps interpretation and responsibility with the human learner.

What we learned

Building The Missing Chart reinforced that safe clinical AI use is not only about whether an answer is correct. It is also about whether the model received the right information, whether important evidence is absent, and whether the user knows which claims require verification.

I also learned that clinicians and AI developers often describe the same problem differently. Clinicians think in terms of evolving evidence, uncertainty, and consequences. Developers often think in terms of inputs, outputs, and system behavior. Effective clinical AI tools must connect both perspectives.

The project also showed that a single interactive workflow can serve several purposes: clinical education, AI-literacy training, evaluation of human-AI interaction, and structured testing of how systems respond to incomplete information.

What's next for The Missing Chart

The next version could include cases from multiple specialties, different patterns of missing information, configurable timelines, and more detailed feedback for learners.

An AI-powered backend could also help educators generate and validate synthetic clinical scenarios, create controlled variations in the information available at each stage, and test how different models respond as the chart evolves. With appropriate clinical review and privacy safeguards, these interactions could produce structured evaluation data for future AI training and safety research.

The long-term goal is broader than one simulation. The Missing Chart could become a shared environment where clinicians learn how to work with AI, developers learn how clinical reasoning actually unfolds, and researchers generate better evidence for building safer medical AI systems. Multiple problems, one connected solution.

Built With

• codex
• gpt-5.6
• next.js
• node.js
• openaisites
• react
• typescript