Predicting Septic Shock Mortality in ICUs with AI

Septic shock, a severe manifestation of sepsis, poses a persistent threat to patients in intensive care units (ICUs), with its complex pathophysiology and high mortality rates making early intervention difficult. Previous machine learning models have attempted to predict sepsis outcomes but were limited by narrow patient scopes, insufficient validation, or low generalisability. Addressing these challenges, researchers developed the TOPSIS-based Classification Fusion (TCF) model, an artificial intelligence solution that integrates multiple algorithms to improve mortality prediction in patients with septic shock. Drawing from over two decades of multicentre ICU data, the TCF model offers promising accuracy, interpretability and clinical value in aiding early decision-making. 

Comprehensive Model Construction and Feature Selection 
The TCF model was trained using data from 4872 patients treated between 2003 and 2023 across three hospitals, encompassing general, paediatric and respiratory ICUs. To build the model, data from 3451 patients were used for training and internal validation, with 1039 patients forming external validation sets. Rigorous preprocessing steps ensured model robustness: features with high missingness were removed, Boolean features with low variance were excluded, and correlation thresholds were applied to minimise redundancy. Missing values were imputed using multiple methods, with logistic regression-based imputation proving most effective. A total of 34 features were ultimately selected based on information entropy and their impact on model performance. These features included vital signs, lab results, treatment history and composite indices such as neutrophil-to-lymphocyte ratio (NLR) and platelet-to-albumin ratio (PAR). 

Seven sub-models—Decision Tree, Random Forest, XGBoost, LightGBM, Naive Bayes, Support Vector Machine and Gradient Boosted Decision Tree—were individually trained before being fused through the Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS). Each sub-model contributed to the final prediction based on weighted performance scores across six metrics: AUC, F1-score, accuracy, precision, sensitivity and specificity. The fusion aimed to balance the strengths and weaknesses of the individual models, ensuring a more stable and accurate final output. 

Model Performance, Interpretation and Clinical Insights 
On internal validation, the TCF model achieved an AUC of 0.733, outperforming all individual sub-models. It also demonstrated superior F1-score, accuracy and precision, indicating consistent and balanced classification ability. While some sub-models excelled in specific metrics, such as LGBM in sensitivity and SVM in specificity, the fusion approach provided the best overall balance. External validation in two independent datasets and two specialised ICUs further confirmed the model’s robustness. Notably, the TCF model achieved an AUC of 0.808 in paediatric ICUs and 0.786 in one external cohort, showing broad applicability across different patient groups. 

Must Read: Enhancing ICU Mortality Prediction with Multimodal Deep Learning 

Interpretability was a key focus of the study. Using SHAP (Shapley Additive Explanations) visualisations for the GBDT sub-model, researchers identified the most influential features in predicting 28-day mortality. These included the number of resuscitations, duration of mechanical ventilation, ICU stay length and diastolic blood pressure. Importantly, all 34 features used were standard clinical variables, enhancing the model’s practicality for routine hospital settings. Weighting feature importance across the sub-models revealed consistent patterns, supporting clinical relevance and transparency in prediction logic. 

Generalisation, Limitations and Future Directions 

The TCF model’s strength lies in its generalisability. Unlike previous models confined to single centres or limited patient groups, this model was validated across three institutions and multiple ICU types. Despite slight performance dips in certain subsets, such as the respiratory ICU, the model maintained robust AUC values and classification capability. A combined validation sample of 1421 patients, including 287 deaths within 28 days, confirmed the model’s effectiveness with an AUC of 0.7705. 

However, the study faced challenges such as data imbalance, missing values and limited positive samples in certain cohorts. These were mitigated through oversampling and advanced imputation techniques. The retrospective nature of the study also limits its ability to confirm real-time clinical utility. Prospective trials across more diverse regions and institutions are recommended to further assess and improve the model’s applicability. Additionally, while the current model relies solely on structured electronic health record data, future versions could integrate unstructured data or multimodal sources to enhance prediction accuracy. 

The TCF model represents a significant advancement in early mortality prediction for septic shock patients. By fusing seven machine learning algorithms into a single, interpretable and generalisable model, it enables more accurate risk stratification across various ICUs. Its foundation in routine clinical data ensures practicality, while its strong performance across multicentre datasets offers confidence in real-world applicability. Models like TCF exemplify how intelligent fusion strategies can elevate predictive healthcare and ultimately contribute to better patient outcomes. 

Source: npj digital medicine

Image Credit: iStock