AI Medical Chronology: Turning Thousands of Pages of Medical Records Into a Reviewable Chronology
is the causation of the injury established ?
Causation of the injury has been established by Dr. Dorman. It has been attributed to the accident
Refer to pages 11-12, 14-15, 21, 180-190
AskGPT can make mistakes. Please check the medical records
Ask a question about the case
I designed an AI Summarization tool that helps Medical Practitioners draft medical chronologies, which will be used by Insurance companies to pay for the plaintiffs affected in Personal Injury Cases. This replaces the process of indexing records, typing out the entire chronology - visit by visit - which will take weeks together. Manual grunt work is removed from their workflow.
23 minutes
saved per medical visit
100%
manual drafting automated
5/5
chronology rating by user
ROLE
Fractional AI Native Product Designer
RESPONSIBILITIES
Own product end to end from research, visual direction, building design system, and hand-off
team
Founder/AI Engineer
What can i help with ?
is the causation established in this case ?
is the causation of the injury established ?
Causation of the injury has been established by Dr. Dorman. It has been attributed to the accident
Refer to pages 11-12, 14-15, 21, 180-190
AskGPT can make mistakes. Please check the medical records
context
Medical Practioners spend days combing through thousands of pages of medical records to manually build chronologies.
personal injury cases
In the USA, a personal injury - accident or medical mal negligence, is followed by medical treatment. The affected person is called Plaintiff. The person or entity causing it is called Defendant.
personal injury cases
In the USA, a personal injury - accident or medical mal negligence, is followed by medical treatment. The affected person is called Plaintiff. The person or entity causing it is called Defendant.
what is medical chronology ?
Post Medical treatment, the electronic medical records are manually drafted into medical chronologies and demand letters. These are used to assess the condition and extent of damage and the Plaintiff will be paid accordingly by the Defendant (his insurance - third party insurance).
weeks consumed in drafting a single chronology
Medical records for each case might range from 100s to 1000s of pages based on the duration of treatment undergone. The practitioners spend days to weeks to draft a single chronology.
Drafting a medical chronology takes about few days to weeks.
Can we build AI Native Product that helps cut short this to hours ?
Can we remove the manual grunt work involved in this ?
Research on AI and LLMS
Turned insights into heuristics for building an AI Native Product
As i was acquainted with the domain, I knew the workflows and did chronologies myself to understand friction points and use cases
I also learnt about LLMs to understand the basics, constraints, and complexities.
Grounding & Trust
The source document should be available to check AI output
Human on the loop
AI drafts; the paralegal decides
Efficiency
The product should let frictionless completion of tasks
Stakes
Since the stakes are high, product direction has been guided by this principles
Product Decisions i owned
Ground AI output in source material
Since the stakes are high, the split screen lets users see the source document and the AI output simultaneously. This way the output can be mapped to source document.
Human on the loop
We have preserved the human agency allowing users to approve, reject, and edit the outputs.
Case centric features
The prototype pushed three different features - chronology and special reports as separate features. However, reports are an adjunct service. Hence the product was made case centric. From the same set of records, user could get the service needed - this way he has to upload records once.
Dashboard
Medical Chronology
User onboarding
I initially designed an empty state, asking users to upload the records. However, this would mean, the user had to wait to see the output (latency) and also spend tokens.
To gain trust on the product, I decided to have a case with an AI output readily available - so a new user can explore the capabilities without having to spend tokens.
Tokenomics & business ROI
Initial push back:
If a feature like special reports is released, the user would use it. However, it is an adjunct service. Whether value from that service will match the token cost for the user is a question. Coz the token cost cannot be shown upfront and user cannot make an informed decision.
How this evolved:
Token costs was discussed and we decided to optimize this.
Skeptical about adoption
What if the users want to continue using humans to index and manually write the chronology ?
What if outsourcing to markets with cheaper labor costs works in their favor ?
What if they do not trust the product since this is high stakes domain ?
Initial Push back - hallucinations, stakes, bias, assumptions, and cognitive surrender
If a feature like AskGPT, ie chat interface is released - the user might over estimate the capabilities of an LLM. Example citations, hallucinations. So is the feature needed in first place is a question. Given the stakes are high, is building one with guardrails justfy the ROI ? Will this scale ? What about cognitive surrender of the user ?
I was initially skeptical of this feature and pushed back.
But eventually found a way to incorporate to align the feature on the same set of principles.
REFLECTIONS
AI cant replace MEEEEE!!!
AI is a misnomer
AI is overrated (people dont understand its limitations)
Guardrails are critical
I am worried about cognitive surrender as a threat to human existence
If you are not complaining about AI slop maybe you were never a thinker in first place
Use AI for code (temperature is low) but not for product strategy - in other words, use it for execution - not strategy
Domain experience is the MOAT
Distribution can be a MOAT
why i picked this case study ?
This was a my first project as founder of my product-design agency. Designing an AI Native Product Ground up meant playing the third wheel between the users and LLMs.
Building a start-up also has the constraint of risk appetite, where you begin to question all product decisions that can affect the business. \
I wanted to show my product thinking, end-to-end ownership, business acumen, and demonstrate how i learnt to design interfaces for LLMs or AI UX - these include non deterministic states, introducing guardrails etc.
Case management tool - enabling workflow rather than AI output
Previously, users were expected to upload merged medical records. This meant, the tool was just a summarizer. However, the tool had to be a case management tool with all records in once place and there is a use case where continuing treatment would mean the chronology be updated. So it is now a case management tool
