Predicting long-term adverse outcomes after neonatal intensive care

This study demonstrates that a time-aware transformer model (STraTS) applied to longitudinal neonatal EHR data can effectively predict long-term neuropsychiatric risks by age seven while providing clinically interpretable insights through multiple complementary interpretability methods, identifying key predictors such as birth weight, Apgar scores, and early clinical severity indicators.

The Gist

Imagine a newborn baby is like a tiny, complex spaceship just launched into the universe. The first 90 days of its life are a chaotic, high-stakes journey through a storm of medical data: blood tests, heart rates, medications, and doctor's notes.

For decades, doctors have known that babies who need intensive care in this "stormy" period are more likely to face long-term challenges later in life, such as autism, epilepsy, or learning difficulties. But predicting exactly which baby will face these challenges has been like trying to guess the weather a year from now by looking at a single snapshot of the sky.

This paper is about building a super-smart weather forecaster for these babies, and more importantly, teaching it how to explain its reasoning so doctors can trust it.

Here is the story of how they did it, broken down into simple parts:

1. The Problem: The "Black Box" Dilemma

Scientists have built powerful AI models (like the one in this study) that can look at a baby's first 90 days of medical records and predict if they might develop a serious condition by age seven.

But there's a catch: AI is often a "Black Box." It gives you an answer ("This baby is high risk"), but it doesn't tell you why. If a doctor can't understand the "why," they can't trust the AI to make life-changing decisions. It's like a GPS telling you to turn left without showing you the map or explaining that the road ahead is blocked.

2. The Solution: The "Time-Aware" Detective

The researchers used a special type of AI called STraTS. Think of most old AI models as a photographer who takes one blurry photo of a baby's entire first 90 days and tries to guess the future from that single image.

STraTS is different. It's more like a detective watching a movie. It doesn't just look at the start and end; it watches the sequence of events. It understands that a fever on day 3 followed by a specific medication on day 5 tells a different story than the same events happening in reverse order. It processes the baby's medical history as a flowing river of time, not a static pile of papers.

3. The Experiment: Testing the Detective

The team fed this AI the medical records of 17,655 children from Helsinki. They asked the AI to predict which of these children would receive a major neuropsychiatric diagnosis by age seven.

• The Result: The AI (STraTS) was the best detective in the room. It outperformed older, simpler models (like Random Forest or Logistic Regression) at spotting the children who were actually at risk.
• The Catch: Even the best AI only got it right about 17% of the time (in terms of a specific metric called AUPRC). This sounds low, but in the world of rare diseases, it's a huge improvement over guessing. It's like finding a needle in a haystack; the AI found more needles than the other tools, but the haystack is still very big.

4. The Real Breakthrough: The "Three-Lens" Glasses

The most important part of this paper isn't just that the AI worked; it's how they proved it was working correctly.

Usually, researchers use just one method to explain why an AI made a decision. The authors realized this is dangerous. It's like trying to understand a 3D object by looking at it through a single pair of glasses. You might see a shadow and think it's a flat circle, when it's actually a sphere.

So, they used three different "lenses" (interpretability methods) to look at the AI's brain:

1. The "What if?" Lens (Perturbation): They asked, "What happens to the prediction if we erase this specific piece of data (like birth weight)?" If the prediction crashes, that data was important.
2. The "Individual Story" Lens (LOO Attribution): They looked at how much each specific piece of data changed the prediction for each individual baby.
3. The "Value" Lens (Value-Dependent Analysis): They checked if higher or lower values (like a higher birth weight) made the risk go up or down.

The Magic Happened When They Compared the Lenses:
By comparing these three views, they found things a single lens would have missed:

• The Consensus: All three lenses agreed on the top 5 risk factors: Birth weight, Gender, Apgar score (how well the baby breathed at birth), Thyroid hormone levels, and how long the baby stayed in the hospital. This gave doctors high confidence that these are real, stable signals.
• The Trap: One lens (LOO) suggested that being born later (higher gestational age) was a risk factor. This sounded wrong! Doctors know that being born earlier (premature) is the risk.
  • Why the confusion? The AI realized that "Birth Weight" and "Gestational Age" are twins; they usually go together. When the AI looked at them separately, it got confused.
  • The Fix: Because they used the other lenses, they saw that "Birth Weight" was the true star, and the "Gestational Age" signal was just a confusing echo. If they had only used one lens, they might have told doctors the wrong thing!

5. The Takeaway: Trust Through Transparency

The study concludes that AI can be a helpful partner in neonatal care, but only if we don't just ask it for an answer. We have to ask it to show its work using multiple methods.

• The Analogy: Imagine a doctor is a captain of a ship. The AI is the radar.
  • In the past, the radar just beeped "Danger!" without showing the screen.
  • In this study, the researchers built a system that shows the radar screen, explains why it beeped, and even cross-checks its own sensors to make sure it isn't seeing a ghost.

In short: This paper shows that by using a "time-aware" AI and checking its logic with three different tools, we can find reliable signals in the chaotic data of a newborn's first 90 days. This helps doctors identify high-risk babies earlier, giving them a head start on care, while ensuring the AI isn't leading them down the wrong path.

Technical Summary

1. Problem Statement

Neonates requiring intensive care are at a significantly elevated risk for long-term neuropsychiatric disorders (e.g., cerebral palsy, epilepsy, autism, intellectual disability). While clinical risk factors are known, predicting individual outcomes remains challenging due to:

• Data Complexity: Neonatal Electronic Health Records (EHR) are irregular, sparse, high-dimensional, and time-dependent, containing heterogeneous data types (diagnoses, labs, medications, vitals).
• Limitations of Current Models: Traditional models often rely on static, time-fixed variables or coarse summaries, failing to capture the complex temporal trajectories of neonatal care.
• The "Black Box" Barrier: Deep learning models (like Transformers) offer superior predictive potential but lack interpretability, hindering clinical adoption. Clinicians need to understand why a model makes a prediction to trust it for decision-making.

Objective: To develop a time-aware deep learning model to predict major neuropsychiatric diagnoses by age seven using the first 90 days of neonatal EHR data, and to rigorously validate the model's predictions using a multi-method interpretability framework to ensure clinical plausibility.

---

2. Methodology

Study Cohort & Data

• Source: Retrospective register-based cohort from Helsinki University Hospital (HUS), Finland.
• Population: 17,655 children admitted to neonatal intensive care or exposed to specific maternal medications during pregnancy.
• Inclusion: Children with at least 7 years of follow-up and valid birth/outcome data.
• Input Window: Longitudinal EHR data from the first 90 days of life.
• Outcome: First occurrence of a major neuropsychiatric diagnosis (Cerebral Palsy, Epilepsy, Intellectual Disability, Autism, Visual/Hearing Impairment) between 90 days and 7 years of age.
• Features: 507 input variables including:
  • 497 time-dependent longitudinal variables (labs, diagnoses, medications, growth, hospital stay duration).
  • 10 time-fixed demographic covariates (gender, birth weight, gestational age, Apgar scores).

Model Architecture

• Primary Model: STraTS (Self-supervised Transformer for Time-Series).
  • Input Representation: Treats data as triplets $(t, r, v)$: time, variable identity, and value. This avoids the need for time discretization or heavy imputation required by dense matrix approaches.
  • Training: Pre-trained via self-supervised forecasting (predicting masked/future values) followed by supervised fine-tuning for binary classification.
  • Mechanism: Uses multi-head self-attention to contextualize sparse, irregular events and attention pooling to generate patient-level embeddings.
• Baselines: Logistic Regression, Random Forest, and XGBoost. These were trained on static, aggregated summaries of the 90-day data (e.g., max/min values, total counts) to ensure a fair comparison.

Interpretability Framework (Key Innovation)

The study employs three complementary methods to avoid the pitfalls of single-method analysis:

1. Perturbation-based Variable Importance:
   • Method: Ablation (removing time-dependent variables) and Permutation (shuffling time-fixed variables).
   • Metric: Change in Area Under the Precision-Recall Curve (AUPRC).
   • Goal: Global ranking of feature importance based on performance degradation.
2. Leave-One-Out (LOO) Feature Attribution:
   • Method: Removing a feature from an individual's input sequence and measuring the change in that specific prediction.
   • Goal: Instance-level explanation and aggregation to a cohort-level ranking. Captures heterogeneity but may underweight rare features.
3. Value-Dependent Effect Analysis:
   • Method: Systematically varying feature values (e.g., quantiles Q10–Q90) while holding others constant to measure the direction and magnitude of risk change.
   • Goal: Translates importance into clinically interpretable effect sizes (e.g., "Higher birth weight reduces risk by X%").

Evaluation Metrics

• Primary: AUPRC (Area Under the Precision-Recall Curve), chosen due to the high class imbalance (8% prevalence).
• Secondary: AUROC (Area Under the Receiver Operating Characteristic Curve).
• Validation: 5-fold cross-validation with strict separation of training, validation, and test sets.

---

3. Key Results

Predictive Performance

• STraTS achieved the highest AUPRC (0.171 ± 0.022), outperforming Random Forest (0.166), Logistic Regression (0.151), and XGBoost (0.128).
• While STraTS had a slightly lower AUROC than baselines, its superior AUPRC indicates better discrimination of the minority class (high-risk infants), which is critical in clinical screening.
• The model successfully identified actionable signals within the first 90 days of life despite the multifactorial nature of the outcomes.

Interpretability Findings

• Convergence: Five predictors were consistently identified as critical across all three interpretability methods:
  1. Birth Weight (Protective: higher weight = lower risk).
  2. Gender (Risk: Male = higher risk).
  3. Apgar Score at 1 minute (Protective: higher score = lower risk).
  4. Umbilical Serum Thyroid Stimulating Hormone (uS-TSH).
  5. Treatment Time in Hospital (Risk: longer stay = higher risk).
• Strongest Risk Indicators:
  • Chromosomal Abnormalities (Q90): Largest effect size (Odds Ratio ~263; +82.9 percentage points risk).
  • Neonatal Cerebral-Status Disturbances (P91): Significant risk increase (OR 3.18).
• Divergence & Artifact Detection:
  • Rare Conditions: Perturbation methods correctly identified rare but severe conditions (e.g., chromosomal abnormalities) as highly important. LOO attribution underweighted these because their effects were diluted when averaged across the whole population.
  • Feature Redundancy Artifact: LOO attribution suggested higher gestational age increased risk (counter-intuitive). This was detected as an artifact because birth weight (highly correlated with gestational age) was already present in the model, capturing the protective signal. Perturbation analysis did not show this error, highlighting the necessity of multi-method validation.

Latent Representations

• PCA visualization of the learned patient embeddings showed that children with similar clinical trajectories clustered together.
• Regions of high predicted risk aligned with established clinical risk dimensions: longer hospital stays, lower birth weight, and lower gestational age.

---

4. Key Contributions

1. Sequence-Based Modeling for Long-Term Outcomes: Demonstrated that Transformer-based models (STraTS) applied to raw, irregular neonatal EHR sequences outperform static summary models in predicting long-term neuropsychiatric outcomes.
2. Multi-Method Interpretability Framework: Established a robust protocol combining Perturbation, LOO Attribution, and Value-Dependent Analysis.
   • This framework successfully distinguished robust clinical signals from method-specific artifacts (e.g., the gestational age redundancy issue) and prevalence biases (e.g., underweighting rare severe conditions in LOO).
3. Clinical Plausibility: Validated that the model's "black box" decisions align with established medical knowledge (e.g., birth weight, Apgar scores, chromosomal abnormalities) while revealing new insights into the predictive power of hospitalization duration and specific lab markers.

---

5. Significance and Limitations

Significance:

• Clinical Utility: Provides a pathway for early risk stratification in neonatal follow-up care, potentially allowing for targeted monitoring of high-risk infants.
• Methodological Advance: The paper argues that in high-dimensional, sparse clinical data, no single interpretability method is sufficient. The combination of methods is essential to build trust and detect hidden biases or artifacts before clinical deployment.
• Data Efficiency: Proves that early-life data (first 90 days) contains sufficient signal for long-term prediction, even without genetic data or later environmental factors.

Limitations:

• Generalizability: The cohort is from a single tertiary center (HUS, Finland); prevalence and predictor distributions may differ elsewhere.
• Data Scope: Excludes genetic data and environmental factors post-90 days, which are known contributors to neuropsychiatric outcomes.
• Outcome Heterogeneity: The composite outcome groups distinct conditions (e.g., autism and cerebral palsy), which may dilute subtype-specific predictors.
• Univariate Focus: The interpretability framework is primarily univariate; complex variable interactions are not fully explored.

Conclusion:
The study demonstrates that combining deep learning with a rigorous, multi-faceted interpretability framework can yield trustworthy, clinically plausible risk predictions for complex long-term outcomes, moving beyond simple accuracy metrics to provide actionable insights for neonatal care.