Ryan F.
McCullough

Systems engineer working on AI early warning for the neonatal ICU. Doctor of Engineering, Penn State.

Ryan F. McCullough
Scatter chart titled Warning Time Is Bought with Candidate Volume. Candidate alerts on the horizontal axis from 4,000 to 11,000, median warning time on the vertical axis from 160 to 320 minutes. Dozens of runs from both architectures lie along one dashed line, 22.65 minutes per 1,000 candidates, R squared 0.963.
Figure 5.3 of the praxis. Each point is one run on the test fold, both architectures, all four training configurations. Another 1,000 candidate alerts buys about 23 more minutes of warning, and the relationship holds across architectures. What differs between designs is where they sit on this line, and the harness is what measures that.

Beyond prediction

A model that predicts late-onset sepsis well is not yet a useful early-warning system. In a neonatal intensive care unit its value depends on what happens after the score: how it is thresholded, how often it is allowed to alert, and how much warning time those alerts buy against the clinician attention they consume.

My doctoral praxis treats that conversion as the engineering problem. It specifies a fixed verification harness in which the alert policy, workload budget, calibration, and metrics are held constant, so that model architecture can be varied as a controlled design variable. Two backbones, PatchTST and a backend-verified Mamba state-space model, were indistinguishable on rank-order discrimination yet produced alert streams that differed by a factor of 2.1 in candidate volume and 110 minutes in median warning time. Repeating one configuration under identical conditions moved warning time by more than the gap between the two models, which is why the harness measures its own run-to-run variation before any difference between designs is reported.

Continuous risk score dashed line: alert threshold Decision logic workload budget refractory interval calibration Delivered alert stream time
After Figure 1.1 of the praxis. The object of study: A model's risk score becomes clinically meaningful only after decision logic turns it into delivered alerts; the harness holds that logic fixed so architectures can be compared on what they cost and what they buy.

The result is a reproducible, claims-bounded method for pricing an AI early-warning model in the units a unit cares about before clinical study: false alerts per patient-day, recall, missed events, and lead time. It is engineering verification on a synthetic cohort, with pipeline feasibility confirmed on authentic neonatal signals from PhysioNet. It is not clinical validation, and the work says so throughout.

A requirements-driven alerting instrument
False alerts per patient-day, refractory behavior, constrained threshold selection, and lead time treated as formal design variables rather than reporting afterthoughts.
A backbone-agnostic self-supervised framework
Representation learning on multichannel neonatal physiology that lets the architecture vary while the evaluation context stays fixed.
Calibration and uncertainty as governance
Both treated as operational requirements of safe thresholding, not optional diagnostics.
A verification ladder
Synthetic-first verification, public-waveform feasibility, and clinical validation kept as distinct levels of evidence with distinct claims.

Publications and talks

About

I am Senior Vice President of Solutions Engineering at a SaaS company whose decision intelligence platform supports planning and prioritization in highly regulated industries. I have been there for over twelve years, most of it spent getting structured decision methods adopted by organizations that have to defend their choices.

I defended my Doctor of Engineering praxis at The Pennsylvania State University in September 2026. The research treats AI for neonatal early warning as a systems-engineering problem rather than a prediction problem: the interesting constraints are alarm burden, patient instability, and finite clinician attention, and the interesting question is how a design behaves inside them.

Working interests: systems engineering for safety-critical health technology, AI verification and validation, workload-aware clinical decision support, physiologic time-series modeling, self-supervised learning, alarm governance, and reproducible engineering evidence packages.

Contact