💻 computer science

Novel Transparent Machine Learning Algorithm - Predicting Sepsis from Routine ED Data.

This retrospective cohort study introduces a novel, fully transparent machine learning algorithm (ShAproxRef) that approximates a "black box" neural network to predict sepsis from routine ED data with comparable accuracy (0.82 AUC on external testing), thereby enhancing clinical trust and facilitating implementation.

Original authors: Alf Timney, Kate Honeyford, Céire Costello

Published 2026-10-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Alf Timney, Kate Honeyford, Céire Costello

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

In the chaotic, high-stakes environment of a hospital emergency department, time is the most critical resource. Every hour that passes before a patient with sepsis receives treatment increases their risk of death by nearly eight percent. Sepsis is a life-threatening reaction to an infection, and while the treatment itself is not complicated, identifying the condition early is notoriously difficult. Doctors rely on algorithms—computer programs designed to spot patterns in patient data—to help flag those at risk. However, the most powerful of these tools often operate as "black boxes." They take in a patient's information and spit out a risk score, but they cannot explain why they reached that conclusion. This lack of transparency makes clinicians hesitant to trust them, as a doctor cannot justify a decision to a patient or a colleague if the reasoning is hidden inside a complex, unreadable code. To solve this, researchers have turned to a method of mathematical fairness called Shapley values, originally developed in game theory to fairly divide profits among players based on their individual contributions. By adapting this concept, scientists can now break down a complex prediction into simple, understandable parts, showing exactly how much each piece of a patient's health record contributed to the final alarm.

A team of researchers from the Institute of Cancer Research, University College London, and Imperial College London set out to build a new kind of sepsis detector that is both accurate and completely transparent. They worked with electronic health records from two National Health Service trusts in the United Kingdom, analyzing data from over 28,000 patient visits. Their goal was to predict whether a patient would eventually be diagnosed with sepsis using only the information available the moment they walked through the emergency room doors. This included routine vital signs like heart rate, temperature, and blood pressure, along with basic demographics and a history of existing health conditions. The team first built a standard, powerful computer model known as a neural network, which is excellent at finding patterns but notoriously difficult to interpret. They then used their novel method to create a second model that mimicked the first one's predictions but did so using a clear, step-by-step logic that any human could follow. This new approach, which they named ShAprox, transformed the hidden calculations of the neural network into a transparent formula where every factor, from age to blood pressure, had a visible and explainable weight.

The results of the study showed that this transparent approach was just as effective as the complex, hidden one. When tested on data from the hospitals where it was trained, the new transparent model correctly identified sepsis cases with a performance score of 0.74, nearly matching the 0.75 score of the original neural network. More importantly, when the team tested the model on data from a completely different hospital to see if it would work in a new environment, the transparent model actually performed slightly better, achieving a score of 0.82 compared to the neural network's 0.81. This suggests that the new method is not only accurate but also robust enough to handle the messy, real-world variations found in different hospitals. The researchers found that the model relied heavily on the most obvious signs of illness, such as temperature and blood pressure, but it also highlighted the importance of less obvious factors. For instance, a patient's age and specific pre-existing conditions, such as lymphoma or liver disease, were significant contributors to the prediction. The model even noted that patients of Pakistani ethnicity appeared less likely to be flagged, a finding the researchers attributed to potential differences in how data was recorded at one of the sites rather than a biological reality, a nuance that would have been impossible to spot without a transparent system.

The true value of this work lies in its ability to bridge the gap between advanced technology and human trust. In a busy emergency room, a doctor needs to know not just that a patient is at risk, but why the computer thinks so. The new algorithm achieves this by assigning a specific contribution score to every piece of data. If a patient has a high fever, the model adds a specific amount to the risk score; if they have a history of cancer, it adds another. This is similar to how a doctor might mentally weigh different symptoms to reach a diagnosis, but the computer does it with mathematical precision and consistency. Unlike other scoring systems that can jump drastically with a tiny change in a single vital sign, this model changes its prediction smoothly, reflecting the gradual nature of how a patient's condition might deteriorate. The researchers also found that they could simplify the model further by removing features that did not add much value, without losing any accuracy. This means the tool could be made even leaner and easier to use in the future.

By proving that a machine learning model can be both highly accurate and fully explainable, this study offers a feasible path forward for implementing artificial intelligence in healthcare. The researchers emphasize that their goal was not to replace doctors, but to provide a tool that supports clinical judgment. Because the logic behind every prediction is visible, clinicians can agree with the alarm, or they can confidently override it if their own experience suggests a different diagnosis. This transparency is essential for adoption; without it, even the most accurate algorithms risk being ignored. The study concludes that by using routine data that is already collected, such as vital signs and basic demographics, hospitals can detect sepsis at the earliest possible moment. This early detection could lead to faster treatment, fewer severe cases, and ultimately, more lives saved, all while keeping the decision-making process clear and understandable for the humans who must act on it.

Technical Summary: Novel Transparent Machine Learning Algorithm for Sepsis Prediction

Problem Statement
Early identification and treatment of sepsis are critical, as every hour of delay increases mortality risk by 7.6%. While the NHS employs digital alerting systems (e.g., NEWS2, St. Johns Sepsis Agent) to detect deterioration, these expert systems often lack the predictive power of machine learning (ML) for early identification prior to clinical decline. Conversely, standard ML models (e.g., neural networks) are often "black boxes." Existing interpretability tools like SHAP (SHapley Additive exPlanations) and LIME provide local explanations but fail to offer a globally transparent algorithm that clinicians can trust and justify. This lack of transparency hinders the implementation of high-performing ML models in clinical settings.

Methodology
The study utilized a retrospective cohort design using electronic patient records from two NHS trusts (Site 1 and Site 2) for patients admitted via the Emergency Department (ED) between June 2019 and August 2020. The dataset comprised 28,072 patient spells, with Site 1 showing a sepsis diagnosis proportion of 10.4% and Site 2 showing 7.0%.

Data Preparation:
- Predictors: Initial vital signs (heart rate, respiratory rate, temperature, systolic blood pressure, Glasgow Coma Scale), demographics (age, gender, ethnicity), and comorbidities (Elixhauser measures, excluding fluid/electrolyte disorders and coagulopathy).
- Handling Missing Data: Missing age/gender led to exclusion. Missing ethnicity was coded as "not known." Missing or extreme vital signs were imputed with random "healthy" values (e.g., heart rate 60–100) using a fixed seed for consistency.
- Outcome: Sepsis diagnosis at any point during the hospital stay, identified via ICD-10 codes.
Model Development:
- NNS1 (Baseline): A shallow neural network (3 hidden layers: 30, 15, 5 neurons; Leaky-ReLU activation; 40% dropout) trained on 55 input variables using TensorFlow. It served as the "black box" benchmark.
- ShAprox (Transparent Approximation): A novel methodology transforming local Shapley value explanations from NNS1 into a globally transparent algorithm.
  - Shapley values were calculated for the training data.
  - Continuous features (vitals, age) were modeled using 3rd-degree polynomial curves of best fit.
  - Binary features were modeled using mean values.
  - The final prediction is the sum of these approximate Shapley values plus a base rate: $F(X) = \sum f_i(x_i) + C$ .
- ShAproxRef (Refined): A simplified version of ShAprox where features with fewer than 50 cases in training or Shapley values between -0.04 and 0.04 were removed.
Evaluation:
- Models were evaluated using Area Under the Receiver Operating Characteristic Curve (ROC AUC) and Sensitivity at Specificity (SAS) at a fixed specificity of 50%.
- Validation included internal testing (20% of Site 1) and external testing (entire Site 2).

Key Contributions

Novel Methodology: This is the first study to transform local Shapley explanations into a globally transparent, additive algorithm. Unlike surrogate models that approximate the black box via regression on inputs, ShAprox approximates the contributions (Shapley values) of features, preserving the parent model's logic while making the decision process explicit.
Transparency without Performance Loss: The study demonstrates that a transparent algorithm can achieve predictive performance comparable to a neural network, addressing the "trust gap" in clinical AI.
Feature Insight: The methodology allows for the identification of key predictive features (e.g., temperature, lymphoma, metastatic cancer) and their specific contribution curves, which is difficult to extract from standard neural networks without retraining.

Results

Performance:
- NNS1 (Black Box): Achieved 0.756 AUC (internal) and 0.818 AUC (external).
- ShAprox: Achieved 0.735 AUC (internal) and 0.824 AUC (external).
- ShAproxRef: Achieved 0.740 AUC (internal) and 0.819 AUC (external).
- The transparent models performed comparably to the neural network on both internal and external datasets. Notably, ShAprox models showed less performance drop between training and testing compared to NNS1, suggesting better generalization.
Feature Importance:
- Temperature was the most predictive feature.
- Vitals: Systolic blood pressure, heart rate, respiratory rate, and GCS score were all in the top 10 most important features.
- Comorbidities: Lymphoma, metastatic cancer, and liver disease were significant predictors (6th, 8th, and 9th most important).
- Demographics: Age was the 5th most important feature. Pakistani ethnicity was the 10th most predictive feature, associated with a reduction in predicted risk (the authors note this may reflect confounding or coding differences rather than causality).
Clinical Metrics: At 50% specificity, the ShAprox models achieved a sensitivity (SAS) of approximately 0.80–0.83, indicating the ability to identify ~80% of sepsis cases with a 50% false positive rate.

Significance and Claims
The paper claims that the ShAprox methodology offers a feasible approach to detect sepsis at the earliest time point (ED triage) using a fully transparent algorithm.

Clinical Trust: By providing a clear, additive equation where clinicians can see how each vital sign and comorbidity contributes to the final probability, the algorithm aims to increase clinician trust and adoption.
Implementation: The model uses only routine data (vitals, demographics, comorbidities) available at triage, avoiding the need for blood tests which may not be immediately available or consistent.
Adjustability: Unlike black-box models, the transparent nature of ShAprox allows for easier iterative adjustment based on clinical feedback (e.g., removing specific features or adjusting thresholds) without retraining a complex neural network.
Generalizability: The study emphasizes the importance of external validation, noting that all models performed better on external data than internal data, highlighting the variability in real-world NHS settings.

The authors conclude that while the ShAprox algorithm matches the predictive power of the neural network, its primary value lies in its high transparency and adjustability, which are essential for clinical implementation and improving patient outcomes through earlier sepsis treatment.

Technical Summary: Novel Transparent Machine Learning Algorithm for Sepsis Prediction

More like this