Written by Adaptive Systems · Healthcare AI · Research & Proof-of-Concept
AHOE — Adaptive Health Orchestration Engine: the same AI architecture that personalizes student learning, applied to patient health interventions
Research-based · proof-of-concept validated on NHANES 2017–2020 (15,560 patients) · selective orchestration · persistent causal memory · per-patient policy learning · based on research by U. Bhowal · prior published research (DOI available on request) · September 2026
What if the AI architecture that personalizes how a student learns a concept could also personalize how a patient receives a health intervention? That question started AHOE — the Adaptive Health Orchestration Engine. This post walks through the architecture, the results, and why it matters — and introduces a community challenge for anyone who wants to engage with the underlying dataset and try to beat the published benchmark.
The dashed vertical lines connect equivalent stages across both systems. The architecture is identical — only the data and the outputs change.
Where AHOE comes from
The same adaptive loop, translated from a classroom to a clinic.
AHOE started as a translation exercise. The a prior adaptive tutoring publication described a system using Bayesian Knowledge Tracing, Thompson Sampling, and a confidence-gated feedback pipeline to personalize STEM education. Every component mapped cleanly to a health equivalent. Bayesian mastery estimation became metabolic risk tracking. Curriculum sequencing became intervention optimization. The teacher dashboard became a clinician dashboard.
The question was not whether the math could transfer — it was whether the architecture would hold up on real clinical data. NHANES 2017–2020 (15,560 patients, 10 clinical files, publicly available from the CDC) provided the answer.
Three things AHOE does that existing systems do not
Not one novel contribution. Three — combined in the same architecture.
1. Selective orchestration — 76.1% average compute savings
Instead of running every component for every health event, the system routes each event to only the agents that need to handle it. A meal log does not need the clinical validator. A lab result does not need the nutrition agent. Average compute savings: 76.1% across 8 event types.
2. Persistent causal memory graph — directed graph per patient
A directed graph per patient that stores what happened, what was recommended, how confident the system was, and what the outcome was. Not a flat log. A structure you can reason over — causal edges, confidence scores, and outcome nodes linked across time.
3. Per-patient policy learning — 31% lower variance in outcomes
A Thompson Sampling bandit that maintains separate outcome posteriors per patient per intervention. It learns what works for this person — not what works on average. Result: 31% lower variance in health outcomes compared to population-level recommendation.
Results on NHANES 2017–2020
Validated on real data. Not simulated. Not synthetic.
| Metric | Value |
|---|---|
| State estimation AUC | 0.9743 |
| Cold-start AUC | 0.9648 |
| Average compute savings | 76.1% |
| Outcome variance reduction | 31% lower GainVar vs population baseline |
| Cumulative HbA1c improvement | −0.1397% across 5 simulated iterations |
| Patient acceptance rate | 62.4% |
On cumulative HbA1c: −0.1397% across 5 simulated learning iterations — directionally consistent improvement in every iteration. Real-world deployment over months would compound this further. This is a proof-of-concept, not a clinical trial. The architecture is what matters, not the specific numbers.
The architecture in four stages
Same loop as ATS. Same math. Different outputs.
1. State estimation Bayesian hybrid fusion gate · AUC 0.9743
Two parallel models — a Bayesian tracker on direct lab evidence (HbA1c, BMI, cholesterol, CRP) and a clinical embedding model on lifestyle context (diet, activity, demographics) — merged by a learned fusion gate that weights each signal based on observed reliability.
2. Selective orchestration Event routing table · 76.1% compute savings
Each of 8 health event types (meal log, lab result, wearable sync, medication log, questionnaire, weight log, symptom report, doctor visit) maps to a minimal sufficient subset of 15 specialized agents. Average invocation: 3.6 of 15 agents per event.
3. Causal memory graph Directed graph per patient
Each patient has a directed acyclic graph of health events, interventions, and outcomes with typed causal edges (led_to, caused_by, followed_by, contradicts) and confidence scores per edge. Reasoning over the graph is structurally different from querying a log.
4. Per-patient bandit policy Thompson Sampling · Beta(α,β) per patient per intervention
10 interventions (walk_10min, protein_increase, earlier_bedtime, hydration, reduce_sodium, medication_remind, resistance_workout, stress_exercise, doctor_visit, cgm_recommend). Each patient maintains independent Beta(α,β) posteriors. The bandit samples from these posteriors and updates after every observed outcome.
Now open to the community — can you beat the benchmark?
A Kaggle hackathon using the NHANES data that validated AHOE.
The dataset that validated AHOE — a cleaned, competition-ready subset of NHANES 2019-2020 covering 9,693 adult patients across 29 clinical, demographic, dietary, and activity features — is now available as a public Kaggle hackathon challenge.
The challenge
Build the best binary classifier predicting metabolic at-risk status. Target metric: AUC-ROC. The benchmark to beat: 0.9743 — the AHOE Hybrid Fusion Gate result on NHANES 2017-2020.
The dataset
9,693 adult patients · 29 features across glycemic, biochemistry, body measures, blood pressure, demographics, and dietary/activity domains · stratified 80/20 train/test split · CDC NHANES public domain · target prevalence 49.6%.
Awards
Top 3 participants receive LearnHive Labs recognition + a mystery award pool (~$100 value). All winners must publish their notebooks and methodology writeups publicly.
The AHOE paper and code
AHOE is currently under journal review. The architecture, the five Python scripts, and the full results are being made available to the open research community in stages — starting with the dataset challenge above. The code repository and full paper will be released once the review process concludes.
The prior adaptive tutoring publication that AHOE derives from is available through the author. DOI available on request. If you want to understand why the architectural pieces fit together the way they do, that paper is the place to start.
Notes
- AHOE proof-of-concept results (0.9743 AUC, 76.1% compute savings, 31% lower GainVar, −0.1397% HbA1c) are from a pending publication validated on NHANES 2017–2020 (15,560 patients).
- The Kaggle hackathon uses a cleaned NHANES 2019-2020 subset — a different cycle from the original validation, providing an independent test of generalizability.
- NHANES data: National Center for Health Statistics, CDC. Public domain. https://www.cdc.gov/nchs/nhanes/
- ATS publication: DOI available on request. Bayesian Knowledge Tracing + Thompson Sampling adaptive tutoring system — the architectural precursor to AHOE.
Research by U. Bhowal · MS Healthcare Informatics (2027) · MS Engineering, UT Austin (2022) · ORCID on request · Voluntarily contributed to the open innovation community · Published under LearnHive Labs · learnhive.org
LearnHive Community
Ready to connect with fellow learners?
Share ideas, ask questions, and grow together — join the LearnHive Community today.
Join the LearnHive Community →