AI in Healthcare: Radiology, Triage, and Clinical NLP Explained

A student in our Spring 2025 AI for Healthcare & Life Sciences cohort spent three weeks trying to get a chest X-ray classifier to stop predicting pneumonia on every single image. The model had 94% accuracy on paper. In practice, it was useless. The problem wasn't the model. It was a subtle label imbalance baked in at the DICOM export stage, and she didn't catch it until we looked at the confusion matrix together in office hours.

That's what AI in healthcare actually looks like. Not a magic diagnostic oracle. A messy, high-stakes engineering problem where the cost of a bug is measured in patient outcomes, not user churn.

Here's what's really happening under the hood in three of the most active areas: radiology AI, emergency triage systems, and clinical NLP.

How Does AI Radiology Actually Work?

AI radiology is a computer vision problem at its core. Most production systems are CNN-based, think DenseNet-121 or EfficientNet variants, trained on labeled DICOM images. CheXNet, the Stanford model that got everyone's attention in 2017, used a DenseNet-121 trained on 112,000 frontal chest X-rays from the NIH ChestX-ray14 dataset. It matched or exceeded average radiologist accuracy across 14 pathology labels.

The hard parts aren't the architecture. They're the data pipeline. DICOM files carry metadata: patient name, date of birth, institution. Every single tag needs scrubbing before the image touches a training environment. We run that de-identification step before data enters our RAG & Vector DB or LLM Fine-Tuning (A100) environments for any healthcare project. One missed tag and you have a HIPAA violation sitting in your training logs.

Beyond classification, the field is moving toward object detection (bounding boxes around nodules) and segmentation (pixel-level tumor delineation). nnU-Net has become close to a default for 3D medical image segmentation because it self-configures based on dataset properties. Radiologists still review every output. The AI surfaces candidates; the human makes the call. That's the design pattern most FDA-cleared tools use, and I think it's the right one.

Windowing surprises a lot of people the first time. A chest X-ray and a CT scan are both grayscale, but the pixel intensity ranges are completely different. You have to apply the correct window-level and window-width settings per modality and per anatomical region. Get that wrong and your model trains on garbage from the start.

Why Clinical NLP Is Harder Than General NLP

Clinical NLP is where I've seen the most overconfident engineering. People take a BERT model, fine-tune it on some medical QA data, and assume it can parse discharge summaries. It can't. Not reliably.

Physician notes are not English. They're abbreviations, shorthand, implied context, negation. The sentence "patient denies chest pain" means the opposite of "patient reports chest pain," but a naive model reading the token sequence can get that wrong in ways that propagate downstream into real clinical decisions. Negation detection, speculation handling ("rule out MI"), and temporal reasoning ("history of" vs. "current presentation") are not solved problems in general NLP. They require domain-specific training data and often rule-based post-processing layered on top of model output.

MIMIC-III is where most teams start. De-identified clinical text from ICU stays at Beth Israel Deaconess. Access is free but gated through PhysioNet credentialing, which takes a few days to a week to clear. Once you have it, a standard pipeline looks roughly like this: extract notes, run a medical NER model (scispaCy or a fine-tuned BioBERT), map entities to SNOMED or ICD-10 codes, then store structured output in a relational layer for downstream analytics.

Fine-tuning a 7B clinical LLM, say a Llama 3.1 7B base with MIMIC fine-tuning, on a single A100 80GB runs roughly 18 to 36 hours depending on sequence length and dataset size. We've walked through that full pipeline in our NLP & LLM Engineering course. The biggest time sink is almost always tokenizer configuration and sequence truncation strategy. Not the training loop itself. That part is usually the easy part.

What Does Real Triage AI Look Like in an ER?

Triage AI does not walk up to a patient and say "you're having a heart attack." It runs as a scoring layer on top of existing EHR data, typically ingesting vitals, chief complaint text, age, past diagnoses, and lab flags to output an acuity score that a nurse or physician reviews.

Systems like EPIC's Deterioration Index, or the models deployed at Vanderbilt University Medical Center, use gradient-boosted trees or lightweight neural networks. Not giant LLMs. Latency matters: you need a prediction in under two seconds, updated every few minutes as new vitals come in. A 70B parameter model is not making that window.

The feature engineering is where the real domain expertise lives. Heart rate divided by systolic blood pressure gives you the shock index. That single derived feature outperforms several raw vitals combined in sepsis prediction. Healthcare ML is full of these clinical knowledge shortcuts that a data scientist without medical context would never derive from raw data alone.

Clinicians also need to override the model. Always. The AI scores a patient as low acuity; the nurse takes one look and escalates. That override data, collected consistently, becomes the most valuable retraining signal you have. Build the feedback loop in from day one. Skip it and you'll be scrambling when the model drifts six months later, and it will drift.

Why Benchmark Performance Lies to You

Every week in our AI for Healthcare & Life Sciences course we talk about what I call the deployment gap. A model hits 96% AUC on a held-out test set and the team celebrates. Then it goes into a real hospital and performance drops to 88% within three months because patient demographics shift, scanning equipment gets upgraded, or documentation practices change after a new EHR rollout.

Distribution shift in healthcare AI is not a theoretical concern. It's routine. Monitoring pipelines are not optional. Our Monitoring & Scaling lab environment exists precisely for this: you need to track prediction confidence distributions, label drift, and downstream clinical outcomes together. Not just model accuracy in isolation.

The engineering is tractable. The stakes just mean you can't skip the parts that feel boring, which is honestly true of most production AI work. Healthcare makes the consequences visible faster than almost any other domain.

If you want to get hands-on with medical imaging pipelines and clinical NLP from first principles, the AI for Healthcare & Life Sciences course is where we do exactly that. Real DICOM data. Real MIMIC text. Real de-identification pipelines before the GPU ever gets involved.