Agreement is often only fair
When doctors judge clubbing by eye, studies found agreement that was often only fair. Kappa values ranged from about 0.36 to 0.90, and several sat near 0.4 [1].
Aarav Shah
I could not stop thinking about a sign you can see on someone's hands, and how easy it is to miss.
I built LungLens because a sign on someone's hands is still judged by a look.
LungLens is an experimental engineering project. It is not a clinically validated diagnostic or screening device.
A clinician looking at the fingers. Photograph for context, not a clinic case.
Why I picked this
Digital clubbing is a change in the fingertips and nails. It can be a reason a doctor looks for lung, heart, or bowel disease.
Someone is holding out their hands. Two trained people look. They may not say the same thing. A parent is listening. I did not want to shrug at that.
I wanted a calmer way to look again, with a photograph. I did not want a machine that claims to know more than it does.
The numbers that stayed with me
A few published facts would not leave me alone. They are from papers, not from LungLens.
The difference can be a few degrees
In people without clubbing, the nail-fold angle usually stays at or below 176°, and doctors are asked to pay attention past 180° [1]. Four degrees on a fingertip is how small the change is, and why a photograph felt worth building.
People already wait too long for some lung answers
In a European study, children with primary ciliary dyskinesia were often not diagnosed until about 5.3 years of age [9]. I am not claiming clubbing caused that delay, but five years is a long time to be unwell without a clear name.
The same review found that a finger depth ratio above 1.0 had a likelihood ratio of 3.9 for lung cancer [1]. That is not a diagnosis. It is a reason someone might need a closer look. These numbers are from papers, not LungLens results. LungLens does not diagnose primary ciliary dyskinesia or any other disease.
How compassion showed up in the build
Compassion, for me, was not a slogan. It was a set of refusals.
Reject the photo
If the photo is blurry, glaring, or badly framed, I reject it. I do not measure a bad picture and call it a finding.
Keep math honest
I keep the angle math separate from any AI that tries to find landmarks. A number should not pretend the anatomy was found.
Say I cannot tell
If I am not confident, I say Unable to assess. I do not turn uncertainty into comforting news.
Do not fake a clinical result
Until the models are actually validated, results stay research-only. A placeholder should not look clinical.
What I can show
The angle math is ordinary geometry. Finding landmarks with AI is still experimental.
- 01 Capture
- 02 Quality
- 03 Find landmarks
- 04 Measure
- 05 Abstain
- 06 Release
17 automated tests passed
I ran engineering checks for the angle math, quality reject and pass behavior, fusion policy, and screening labels (2026-10-19).
Not clinically validated
I have not proven clinical accuracy. Finding landmarks with AI is still experimental. Passing software tests does not prove the system works across lighting, cameras, or skin tones.