Darrell A. Hall, MD, MSCIS

Most healthcare machine learning fails at the join between the clinic and the model — not because the model is weak, but because nobody in the room can see both sides at once. I have practiced medicine and I have built systems. I work on the problems that need both.

What I do

• Clinical validation of ML systems. Study design, operating-point selection, error budgets, and the question most models never answer: what does this system do when its assumptions are false?
• Spatial and population epidemiology. Bayesian spatiotemporal modeling, county and census-tract analysis, and careful handling of the suppression and small-count problems that quietly invalidate public-health findings.
• Production clinical systems. Postgres schema design, de-identification architecture, role-based access control, and the statutory obligations that shape both.
• Device and diagnostic modeling. Physics-grounded simulation and sensor-chain analysis for systems that have not been built yet.

Selected work

NABS — thermographic pressure-injury surveillance

device modeling · validation simulation · signal detection

A hygiene enclosure for non-ambulatory patients that also performs long-wave infrared surveillance of eight bony prominences. I built the physics, the scheduler, the detector and the validation simulator.

The simulator is anchored to a published diagnostic-accuracy meta-analysis rather than to an assumed accuracy, and it produced a result that reversed the design: instrument noise is under 5% of the error variance, while physiological reference variance is over half. The constraint was never the camera. It was that ΔT has no meaning until you specify what is being subtracted.

Interactive demo Full results

County-level overdose mortality — the 2022–24 reversal

Bayesian spatiotemporal modeling · INLA · in progress

A national study of the reversal in overdose mortality, using BYM/CAR spatiotemporal models fitted in INLA across a county panel. The central methodological problem is CDC WONDER suppression: small-count counties are withheld, and the withholding is not random with respect to the outcome. Any analysis that treats suppressed cells as missing-at-random will produce a confident and wrong answer about precisely the rural counties the question is about.

Structured as a book with two to three peer-reviewed papers beneath it, paired with a document retrieval corpus over the opioid industry archives.

GeoCode Analytics — substance use risk mapping

geospatial analysis · R · Lucas County, Ohio

Census-tract-level mapping of substance use risk across Lucas County, built in R with tidygeocoder and Leaflet over a cleaned address corpus of roughly 7,200 records, with eight-sector directional analysis and a path for wearable telemetry integration.

KATA — youth behavioral health records

production system · FastAPI · Postgres · Cloud Run

A behavioral health record and safety-workflow system for programs serving minors, built against Ohio statutory obligations including mandated reporting under ORC 2151.421. Postgres schema with a pgcrypto de-identification view and a separate analyst role, Pydantic contracts, FastAPI with six-role RBAC, Alembic migrations, GitHub Actions CI with 51 passing tests, and a React/TypeScript frontend split into clinician and de-identified analyst dashboards. Deployed on Cloud Run and Cloud SQL.

The interesting part was not the API. It was designing the schema so that the de-identified analytic view and the statutory reporting obligation could both be satisfied without either one compromising the other.

PMSG — non-pharmacologic chronic pain program

clinical program design · retrieval platform · R Shiny analytics

An eight-week, 32-session, 80-contact-hour chronic pain program integrating pain neuroscience education, CBT, anti-inflammatory nutrition and movement modalities — with a full platform behind it: a Postgres schema of 20+ tables, a ChromaDB/LangChain retrieval pipeline, a React frontend and R Shiny analytics, documented in an 18-chapter technical manual.

Inner Combat Training

community program · intake and vetting on GCP

A free youth martial arts and behavioral wellness program in Toledo serving ages 6–16 across four cohorts, with intake and vetting workflows running on Cloud Run and Cloud SQL. Awarded a community wellness grant in 2026.

Background

I trained in family medicine and practiced for fourteen years, much of it with non-ambulatory patients. Before medicine I spent seven years as a United States Air Force communications-electronics officer at European Command, working on packet-switched network programs including the Defense Data Network — which is where I learned what happens when a system reports a confident number it has no basis for.

I hold an MSCIS from Boston University, where I took my first artificial intelligence course in 1985. I returned to the field in 2017 and certified as a Professional Data Scientist in 2022.

The combination is unusual and it is the point. Clinically I know what a measurement has to mean before it is worth anything. From the engineering side I know how to build the thing that produces it.

Working together

I take engagements in clinical validation, spatial and population epidemiology, applied machine learning, and production clinical systems — as a consultant, a technical advisor, or a second pair of eyes on work already underway. Rates depend on scope; I am glad to talk it through.

I am comfortable being the person who tells you the model is wrong.