📄 medicine

PORT: Generative EHR Pretraining for Pediatric Intraoperative Risk Prediction

The paper introduces PORT, a generative transformer model pretrained on pediatric EHR timelines that, when adapted with low-rank fine-tuning, achieves high accuracy and calibration in predicting rare intraoperative deterioration events from pre-operative data, thereby enabling proactive risk stratification.

Original authors: Junseob Kim, Haoming Shi, Michael C. Fiedorek, Hannah D. Kilday, Henry P. Foote, Christoph P. Hornik, Rishikesan Kamaleswaran, Justin B. Long

Published 2026-08-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Junseob Kim, Haoming Shi, Michael C. Fiedorek, Hannah D. Kilday, Henry P. Foote, Christoph P. Hornik, Rishikesan Kamaleswaran, Justin B. Long

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery before it even happens. In the world of medicine, this is called risk prediction. Doctors have always tried to guess which patients might get into trouble during surgery, but for children, this is especially tricky. The events they are looking for—like a sudden heart problem or a crash in blood pressure—are rare, like finding a single specific grain of sand on a massive beach. Traditional methods are like using a magnifying glass to look at just a few grains of sand; they might miss the big picture.

To solve this, scientists have started using "generative pretraining." Think of this as teaching a robot to read millions of medical stories (Electronic Health Records) just to understand how the human body works, without asking it to solve a specific puzzle yet. It's like letting a student read every textbook in a library before asking them to take a test. Once the robot has learned the "language" of medicine, scientists can give it a tiny, lightweight upgrade (called "Low-Rank Adaptation") to teach it how to spot a specific danger. This paper asks: Can we use this powerful, pre-trained robot brain to predict rare, scary moments in children's surgeries before they even happen?

The researchers built a new tool called PORT (Pediatric Operative Risk Transformer). They trained it on the medical histories of nearly 190,000 anesthesia encounters from children at Children's Healthcare of Atlanta between 2014 and 2021. The goal was to predict Intraoperative Deterioration (IoD), which is a fancy way of saying "something goes wrong during the surgery that needs immediate, serious help," like a heart stopping or needing emergency medicine.

Here is how PORT works and what it found:

The Magic of the "Pre-trained Brain"
Instead of building a model from scratch, the team first taught a "backbone" model to predict the next event in a patient's medical timeline. Imagine a robot reading a story about a patient's life, learning that a fever often leads to a hospital visit, or that a specific heart condition usually requires a certain medication. This robot learned the patterns of 127,874 different patients. Then, the team gave this robot a "specialized hat" (the LoRA adapter) to focus specifically on spotting the signs of a surgery disaster.

The Results: A Super-Detective
When they tested PORT on a group of patients it had never seen before, it was incredibly good at its job.

Accuracy: It correctly identified the risk of a bad event 94.2% of the time (an AUROC of 0.942). This was significantly better than the next-best model, which got about 90.6% right.
Calibration: It didn't just guess; it was honest about its confidence. If it said there was a 10% chance of trouble, trouble happened about 10% of the time. This is crucial because doctors need to trust the numbers.
The "Needle in a Haystack" Problem: Since these bad events are rare (happening in only 0.27% of cases), most models get confused. PORT, however, managed to catch 31.7% of all the bad events while only flagging 0.4% of the patients as "high risk." This means it could help doctors focus their extra attention on a tiny group of kids who really need it, rather than worrying about everyone.

Why It's Different
Older models were like checking a checklist: "Does the patient have a fever? Yes. Do they have a heart condition? Yes." PORT is more like reading the whole story. It looked at the entire timeline of a child's life before surgery, including their past surgeries, medications, lab results, and even how long it had been since their last visit.

The "Time Travel" Test: The researchers tested if PORT would still work on data from the future (patients seen after 2019). It did! It maintained its high accuracy, suggesting it learned real medical patterns rather than just memorizing old data.
The "Why" Factor: The paper didn't just say "it works"; it explained why. Using a technique called Integrated Gradients, they could see which parts of the medical history made the robot nervous. For example, if a child was on heart-strengthening drugs or had a history of being in the ICU, the model flagged that as a major risk. This helps doctors understand the prediction, not just trust a black box.

What It's Not
The paper is careful to say what PORT is not. It is not a crystal ball that works perfectly for every single child.

Heart Surgery: The model worked best for general surgeries. For children with complex, congenital heart defects, the model was less accurate. The authors suggest this is because these patients are already so high-risk that the "extra" signals PORT looks for are less helpful; for them, specialized heart surgery scores are still better.
Not a Replacement: PORT is a tool to help doctors plan, not a replacement for a doctor's judgment. It suggests who might need extra monitoring or a specialist team standing by, but it doesn't perform the surgery.

The Bottom Line
This paper suggests that by teaching a computer to read the "language" of medical records first, and then giving it a tiny, efficient upgrade, we can build a system that spots rare, dangerous events in children's surgeries with high accuracy and honesty. It's a step toward a future where doctors can say, "This child is high-risk, let's get the extra team ready," before the surgery even begins, potentially preventing disasters before they happen. The authors note that while the results are promising, the tool still needs to be tested in real-time, live surgeries before it becomes a standard part of hospital care.

Technical Summary: PORT (Pediatric Operative Risk Transformer)

Problem Statement

Pediatric intraoperative deterioration (IoD) represents a rare but high-consequence event, including hemodynamic collapse, cardiac arrest, and unplanned vasoactive escalation. While early identification of high-risk patients could facilitate proactive perioperative planning (e.g., specialist coverage, ECMO standby), current risk-stratification tools are limited. Existing approaches rely on manually curated features (e.g., ASA Physical Status, RACHS-1, STS-CHSD) or coarse perioperative summaries, which often lack inter-rater reliability and fail to capture the complex temporal structure of longitudinal electronic health record (EHR) data. Furthermore, machine learning models in pediatric anesthesia have historically depended on engineered features and single point estimates without quantifying predictive uncertainty or providing interpretable, patient-level explanations.

Methodology

Data and Cohort

The study utilized a retrospective, single-center cohort from Children's Healthcare of Atlanta (CHOA) comprising 189,704 anesthesia encounters from 127,874 patients (2014–2021). Data were harmonized from 12 structured EHR tables into the Medical Event Data Standard (MEDS) format, resulting in 26.5 million clinical events across 53,705 unique codes. The dataset was split at the patient level (70% training, 10% validation, 20% test) to prevent data leakage, with a temporally held-out post-2019 cohort used for robustness testing.

Outcome Definition

IoD was defined as a composite of five intraprocedure events requiring active intervention:

Cardiac arrest requiring CPR.
Free-text deterioration identified via keyword matching, LLM screening, and clinician adjudication in anesthesia "Quick Notes."
IV vasoactive bolus (e.g., epinephrine, phenylephrine) with specific dose thresholds.
Vasoactive infusion escalation (>100% rate increase).
Emergent arterial line placement after surgical incision.

Model Architecture: PORT

PORT (Pediatric Operative Risk Transformer) is a decoder-only transformer architecture adapted from the ETHOS framework.

Tokenization: Heterogeneous EHR timelines were tokenized into a 6,170-token vocabulary. Numeric values (labs, vitals, doses) were paired with quantile tokens, and time gaps were discretized into 19 interval tokens. The vocabulary preserves code identity while handling long-tail distributions via "OTHER" bins.
Pretraining: The backbone (a 6-layer, 47M parameter GPT-2 style model) was pretrained autoregressively on the full MEDS timeline using next-token prediction. This was trained from random initialization on pediatric surgical data, distinct from adult corpora.
Adaptation (LoRA): To predict IoD, the pretrained backbone was frozen, and Low-Rank Adaptation (LoRA) adapters were inserted into the query-key-value (QKV) projections of all six attention layers. A lightweight 2-layer MLP classification head mapped the hidden state at the "OR entry" token to the IoD probability. This approach trained only 0.73% of the total parameters.
Uncertainty Estimation: Predictive uncertainty was quantified by ensembling three independently initialized models and measuring the standard deviation of their predictions.

Baselines

The model was compared against:

Clinical Baseline: ASA Physical Status score.
Tabular Models: Logistic Regression (LR) and XGBoost using manual features and MEDS-aggregated summaries.
Sequence Models: A tuned bidirectional LSTM (BiLSTM) trained on raw MEDS sequences.
Zero-Shot Inference: The pretrained backbone used without supervised fine-tuning, estimating probability via trajectory sampling.

Key Results

Discrimination and Calibration

On the held-out test set (37,948 encounters, 101 IoD events), PORT achieved:

AUROC: 0.942 (95% CI: 0.915–0.964), significantly outperforming the best baseline BiLSTM (0.906, $p=0.006$ ).
AUPRC: 0.237, compared to 0.191 for the BiLSTM.
Calibration: The model was well-calibrated with an Expected Calibration Error (ECE) of 0.001, requiring no auxiliary calibration.
Zero-Shot vs. Supervised: Zero-shot trajectory inference performed poorly (AUROC 0.588), confirming that supervised adaptation is necessary for composite outcomes involving unstructured text components.

Temporal Generalizability

In a temporally held-out test (patients first seen post-2019), PORT maintained strong discrimination (AUROC 0.949), outperforming all baselines. While AUPRC decreased (0.051) due to lower event prevalence (0.11%) and shorter pre-operative histories, the model remained superior in ranking and calibration.

Interpretability and Attribution

Integrated Gradients (IG): Per-encounter attribution identified clinically salient drivers. For high-risk cases, risk was driven by active cardiovascular support (vasopressors, inotropes) and ICU admission.
Category Occlusion: Masking "Surgical Context" tokens (procedure, ASA, admission type) caused the largest drop in AUROC ( $\Delta$ AUROC = -0.056). However, masking "Medications" caused the largest drop in AUPRC ( $\Delta$ AUPRC = -0.189), indicating that medication intensity (e.g., vasoactive use) is critical for identifying rare positive events.
Risk Stratification: Flagging the top 10% of risk scores captured 83.2% of IoD events, compared to 72.4% for the BiLSTM. A high-risk subgroup (>10% predicted probability) comprising only 0.4% of encounters captured 31.7% of all IoD events.

Efficiency

Inference required only 7.2 ms per encounter and 318 MB of GPU memory. Performance remained robust using only a 7-day pre-operative history window (AUROC 0.943), suggesting feasibility for institutions with limited longitudinal data.

Significance and Claims

The paper claims that PORT represents the first application of generative EHR pretraining to pediatric intraoperative risk prediction. Its primary contributions are:

A Reusable Foundation: It establishes a pretrained generative backbone that can be efficiently adapted to related perioperative outcomes via lightweight LoRA adapters, moving beyond task-specific supervised learning.
Clinical Utility: It provides calibrated, uncertainty-aware risk estimates that support proactive planning (e.g., staffing, resource allocation) without requiring post-hoc recalibration.
Interpretability: The model grounds predictions in specific, named perioperative tokens, enabling case-level audits and revealing that elevated risk often reflects baseline acuity (e.g., chronic medication use) rather than just acute instability.
Data Efficiency: Despite being trained on a specialized pediatric cohort with sparse rare-event labels, the generative pretraining approach yielded superior performance compared to models trained from scratch, provided sufficient training data (approx. 50% of the cohort) was available.

The authors conclude that while PORT is not a replacement for procedure-specific scores in high-acuity cardiac surgery (where discrimination was limited for all models), it offers a scalable, interpretable framework for risk stratification in the broader pediatric surgical population, particularly for noncardiac procedures where current tools are less effective.