Patient information in one place. Clear evidence. The clinician decides.
A working patient workspace and two research studies.
The problem
Vital signs, surgery records and test results sit in different systems. Spotting who will need intensive support is hard.
needed organ support (breathing, kidney or heart) or ended in death within 3 days of surgery.
In our study data.Research study · Risk after surgery
Critical organ support or death during the following three days
Recorded critical organ support or death.
We tested the models on patients excluded from training, then compared their predictions with what actually happened.
9 times out of 10, the model ranked the right patient as higher risk.
Given one patient who later needed organ support or died and one who did not, the model gave the higher risk to the right patient about 9 times out of 10. Adding measurements from during and just after surgery raised this from about 8.6 to 9.
This compares one case with organ support or death and one without the outcome. It is not 90% accuracy in predicting death.
AUROC 0.86 → 0.89 → 0.90 · 0.5 = chance, 1.0 = perfect · 95% intervals shown.
Average precision 0.070 vs 0.0018 chance, averaged across risk thresholds. Not an individual patient's probability. The last gain is not statistically confirmed.
A patient’s operations always stay in the same group.
INSPIRE via PhysioNet · Seoul National University Hospital · retrospective
No peeking aheadOnly data available at 6 h was used.
Patients kept separateOne patient’s operations never split across train and test.
Rare events measured honestlyA measure built for rare outcomes, with 95% intervals.
Limits statedSingle hospital; external validation next.
Uses pre-surgery information to predict improvement in pain and daily function after knee replacement.
Early research model, small test. Larger testing next.
Overall study population · Source: NIH-sponsored Osteoarthritis Initiative.
Trained on 98 replacement cases. Tested on 25 separate cases, without showing the model their recovery outcomes.
Knee replacement, left knee
| Value | Read by model | Source variableOriginal code in the research dataset. |
|---|---|---|
| KOOS painknee symptom & function score | Yes | V06KOOSKPLPublic research · OAI. Patient-reported pain measure; read by the recovery model. |
| WOMAC totalarthritis pain & function score | Yes | V06WOMTSLPublic research · OAI. Patient-reported total measure; read by the recovery model. |
| Walking pace | No | V0620MPACEPublic research · OAI clinic visit. Contextual functional information; not read by the recovery model. |
| Kellgren-Lawrence grade | No | V05XRKLPublic research · OAI central X-ray reading. Structured reading, not an image file or model input. |
| Laboratory results | No | —Not yet connected. No laboratory value is supplied. |
Research prototype · Osteoarthritis Initiative research data · demonstration labels · not for clinical use.
KOOS pain links to V06KOOSKPL; WOMAC total links to V06WOMTSL in the OAI research dataset.
Walking pace and X-ray readings are marked as context, not inputs to the recovery model.
Laboratory results, biomarkers and allergies are marked as not yet connected.
The next milestone
Founder & CEO · AI Systems Architect
More than 15 years of social-impact entrepreneurship, leadership and product development, with hands-on experience in AI systems architecture and software development.
My work connects AI and health technology. At Neomage, I develop methods for integrating health data to understand risk, disease progression and recovery, with an emphasis on clinical validity.
My earlier work included educational support and workplace stress-management education for healthcare professionals.
Bring a clinical question. We bring the prototype, the models and a plan to test them on your data.
Talk about a pilotWe're raising pre-seed to build the platform that connects every type of patient data.
Talk about investment