Skip to main content

KOI — AI for Anesthesia | ASA-PS Classification -89% Error | HTX

On-premise AI Act compliant Research Use Only

AI for anesthesia.

Automatic ASA-PS classification with advanced reasoning models. -89% classification error compared to physicians. Clinical data that never leaves your hospital.

-89%
Classification
error reduction
97.5%
Advanced model
accuracy
100%
Clinical data
on-premise
How it works

Decision support for the anesthesiologist.

ASA-PS classification is the standard for preoperative anesthetic risk assessment. But inter-rater variability is high: physicians agree on the correct classification in only 7 out of 10 cases. KOI reduces this variability with advanced reasoning AI.

01

Patient record analysis

KOI receives patient data: medical history, physical examination, diagnostic tests. The AI model analyzes the complete clinical picture.

Chain-of-thought
02

ASA-PS classification

The model produces an ASA-PS classification with detailed clinical reasoning and confidence score. Every decision is explainable and verifiable.

97.5% accuracy
03

Physician decision

The anesthesiologist reviews the proposed classification, compares it with their own clinical judgment, and decides. KOI supports, never replaces.

Human-in-the-loop
Scientific research

Validated on 20 standardized clinical cases.

The HTX study evaluated 11 language models — from early-generation models (GPT-4, LLaMA) to advanced reasoning models (GPT-o3, Claude Sonnet, DeepSeek R1) — on 20 clinical vignettes from the most studied ASA-PS benchmarks in the literature.

Key findings: advanced models achieve 97.5% accuracy (95% CI: 92.9%-99.1%), significantly outperforming both early-generation models (~77%) and the human benchmark (7.7/10). Mean error drops from 2.3 to 0.25 misclassifications per 10 cases.

DeepSeek R1, deployable on-premise, showed perfect reproducibility across repeated trials — demonstrating that privacy-preserving deployment is feasible without sacrificing accuracy.

Peer-reviewed (in submission) Multilingual EN/IT 11 LLMs tested
KOI — ASA-PS Project
KOI
Architecture

Clinical data never leaves the hospital.

KOI runs on PRISMA, HTX's private AI infrastructure. On-premise or European cloud deployment, with end-to-end encryption. No patient data is sent to external servers.

38.4% of healthcare LLM studies fail to implement adequate data protections. KOI is designed for GDPR and AI Act compliance by design — because in healthcare, privacy is not optional.

GDPR EU AI Act ISO 13485 IEC 62304

Want to bring KOI to your hospital?

Contact us for a demo. We'll show you how KOI can support your anesthesia department.

Get in touch

Funded research project

KOI was born from the "ASA-PS Classification" project, funded by the Friuli Venezia Giulia Region (LR 22/2022, art. 7 — TRL 6-8 validation support, EUR 90k grant). In collaboration with Ospedale del Quadrante (Ramsay Sante), the project clinically validates the AI system for ASA-PS classification. Period: December 2025 - November 2026. Currently Research Use Only (RUO) — medical device clearance planned for 2027.

Frequently asked questions

KOI FAQ

What is KOI and how does it work?

KOI automatically classifies patients according to the ASA-PS scale for anesthesiology — 89% fewer errors than manual classification. It uses advanced reasoning models that analyze the patient record and produce a classification with detailed clinical explanation.

Can I deploy KOI in my hospital?

Yes. KOI runs on PRISMA, deployable on-premise. Clinical data never leaves the facility. Full GDPR and AI Act compliance.

Is KOI a medical device?

KOI is currently classified as Research Use Only (RUO) — available for research and clinical validation, not for diagnostic practice. HTX is pursuing ISO 13485 and IEC 62304 certification, with medical device clearance planned for 2027. KOI is developed with Ospedale del Quadrante (Ramsay Sante), supported by a EUR 90,000 grant from the FVG Region.