We test healthcare AI with universities before it reaches membersSince 2025, graduate teams at Northeastern University Vancouver and the University of Victoria have tested retrieval, safety scoring, document security and clinical prediction for us on research data, never member data.
01Research Collaboration
Graduate programs at two universities run these projects under faculty supervision, present them at public showcases, and never use member data.
Khoury College of Computer Sciences and College of Engineering
Graduate capstones in retrieval-augmented treatment planning (2025), ECHNA safety scoring (2026) and document-upload security (2026), with two further tracks under way in Fall 2026.
Since 2025 · 5 projects
Faculty of Engineering and Computer Science, Biomedical Systems
A 2025 graduate project on explainable multi-organ deterioration prediction from hospital records, co-supervised with our founder & CTO, Dean Shev, as industry advisor.
Since 2025 · 1 project
VHT's ECHNA demonstrates a compelling proof-of-concept for ethically-grounded, empathetic AI in healthcare navigation — integrating patient-centered values with rigorous clinical standards.
Research & Capstone ProjectsNortheastern University — VancouverA completed graduate study at UVic evaluated ECHNA's lab-result analysis capabilities, finding that the platform accurately contextualized complex biomarker data and surfaced clinically relevant insights in plain language for patients.
Graduate CollaborationUniversity of Victoria02Faculty collaborators
These faculty members supervised or examined each project, and students are credited on the project cards exactly as printed on the public posters.
Khoury College · Northeastern Vancouver
College of Engineering · Northeastern Vancouver
Khoury College · Northeastern Vancouver
Khoury College · Northeastern Vancouver
CAMTEC · University of Victoria
03Projects
Each project tested one part of healthcare AI on research data and reports one headline result with its limits, linking only to public material.
6 of 6 projects shown
Northeastern University Vancouver · Capstone, February to April 2025
0.796BERTScore F1, plan quality
Can a retrieval-augmented stack find similar patient cases and PubMed literature, then draft a structured, evidence-aware treatment plan with measurable retrieval quality at every layer?
How it was built
Treatment-plan AI
A layered RAG pipeline: vector stores over roughly 512,000 de-identified patient-like records and about 425,000 PubMed documents, similarity retrieval at each layer, and a language model that drafts a plan from the retrieved evidence.
Plan quality averaged BERTScore F1 0.796 (best 0.873) against reference plans; patient-case retrieval reached Precision@10 of 1.0 on all ten test cases.
Evaluation on research data by the student team; not a measure of the VHT product.
Northeastern University Vancouver · Capstone, February to April 2026
4 axesEthics, privacy, autonomy and accuracy, scored per answer
Can a panel of judge models score a medical AI answer on ethics, privacy, autonomy and accuracy, then guide a retrieval-based revision before it reaches a member?
How it was built
LLM Tribunal
A three-judge scoring layer built on open-weight medical and general models, a 50-example expert-annotated set, and a rule corpus of 20 rules and 143 documents drawn from WHO, UNESCO and NIST AI RMF guidance that drives the revision step.
On the showcase question, ECHNA’s answer rose from 32/40 to 34/40 after Tribunal revision, with the privacy score moving from 6 to 8.
32 → 34 / 40Showcase example: rubric score after Tribunal revision
Scores on the team’s own rubric, measured on research prompts; the student code is not part of the live ECHNA service.
How this rubric shapes ECHNA
Northeastern University Vancouver · Research capstone, February to April 2026
0.9996AUC-ROC, PDF malware (LightGBM)
Can a local, HIPAA-compatible pipeline catch both classic PDF malware and prompt-injection text hidden in uploaded medical documents, without sending anything to cloud scanners?
How it was built
Upload defence
A hybrid detector: a LightGBM classifier over structural PDF features for malware, and a locally hosted 8B language model that screens document text for injected instructions, packaged to run inside the platform boundary.
LightGBM reached 99.13% accuracy and 99.96% AUC-ROC (0.74% false-positive rate) on the public CIC-Evasive-PDFMal2022 set; the local model flagged 91% of injected documents at 1.36 s mean latency.
Public datasets and synthetic injections only, static analysis only; about 28% of benign documents were flagged for review by design. Not a product guarantee.
University of Victoria · Graduate project, 2025
0.792–0.876AUROC across six tasks
Can one explainable multi-task transformer predict renal, cardiac, metabolic and perfusion deterioration jointly from hospital event sequences, fast enough for real-time use?
How it was built
Deterioration prediction
A multi-task transformer trained on the MIMIC-IV research database (about 100,000 admissions), with shared representations across six deterioration tasks and built-in explanations, co-supervised with our founder & CTO, Dean Shev, as industry advisor.
AUROC between 0.792 and 0.876 across the six tasks, with under 50 ms inference and 87% fewer parameters than eight single-task models.
Credentialed research data (MIMIC-IV), not VHT member data; nothing from this project runs in the VHT product.
Northeastern University Vancouver · Prof. Maryam Tanha, Supervising faculty
Northeastern University Vancouver · Dr. Hazra Imran, Supervising faculty
04Work with us
If you teach or study at a university, this is how to propose a capstone or research project with VHT.
We host graduate capstones and research projects each term: scoped questions, public or synthetic data, faculty supervision, and a public showcase at the end. Faculty, students and research groups are welcome to propose a project.
ECHNA provides educational health guidance only and does not diagnose or treat. Nothing on this page is FDA-cleared, and none of these research projects is a shipped VHT feature. All results were measured on public or credentialed research data, never on VHT member data.
News & Events
Every talk, showcase, pitch and podcast is now on our News & Events page, alongside company news.
14 talks, showcases, pitches and sponsorships · 5 press features
