Writing

Notes on building clinical AI — from data pipelines to regulatory submissions.

2026-04·8 min·draft

Building Deep Learning Models for FDA Submission

What Changes When the Output Matters

Most ML tutorials end at "the model achieves 0.92 AUC." In clinical AI, that's roughly where the real work begins. Lessons from building an ECG algorithm that received FDA Breakthrough Device Designation.

regulatorydeep-learningFDAclinical-ai
2026-04·7 min·draft

The Data Pipeline Nobody Talks About

Why 80% of Clinical AI Is Data Work

Every clinical AI paper describes the model. Almost none describe how the data got there. On building reproducible pipelines for some of the messiest data in machine learning.

data-engineeringclinical-aipipelines
2026-04·9 min·draft

From CNN to Vision Transformer: Choosing Architectures for ECG Analysis

What the Benchmarks Won't Tell You

I trained and compared CNNs, ResNets, Vision Transformers, CNN-LSTM hybrids, and multimodal models for cardiac disease detection. Here's what works, what doesn't, and why architecture comparisons are harder than they look.

deep-learningarchitecturesECGtransformers
2026-04·8 min·draft

Why Model Validation in Healthcare Is Different From Everything Else

Subgroups, Calibration, and the Evidence Gap

In standard ML, validation means a held-out test set. In clinical AI, it's a multi-month process that determines whether your model reaches a single patient. On building validation frameworks for regulated medical AI.

validationclinical-airegulatorycalibration
2026-04·8 min·draft

FDA vs. CE Mark: Navigating Two Regulatory Worlds for Medical AI

Same Goal, Different Philosophies

I've worked on algorithms that went through both the US and European regulatory pathways. They're often mentioned together as if they're the same. They're not.

regulatoryFDACE-markSaMD
2026-04·10 min·draft

What 5 Years of Garmin Data Taught Me About Training

A Personal Biosignal Analysis Project

HR, HRV, sleep, training load — applying the same signal processing and ML techniques I use professionally to my own wearable data.

wearablesbiosignalsdata-analysisgarminpersonal