Table 1 Study details, model development and performance characteristics of studies predicting functional outcome following moderate to severe TBI at ≥3 months from injury
First author and year of publication | Study-specific TBI definition | Functional outcome assessment time | Outcome categorization | Sample size (adult, mixed, or pediatric) | AI model architecture | AI performance (top model performance reported) | Non-AI performance (top model; if existing nomogram or human experts included, all reported) |
|---|---|---|---|---|---|---|---|
Jung 2023 | Undergoing ICP monitoring in intensive care unit (ICU) | 6 months | Functional outcome (GOS-E 1–3 vs. 4–8) | 166 (adult) | Light Gradient Boosting Machine (GBM) | AUC: 0.858 | Not included |
Gravesteijn 2020 | GCS ≤ 12 | 6 months | Functional outcome (GOS 1–3 vs. 4–5 or GOS-E 1–4 vs. 5–8) | 12,397 (adult) | Support vector machine (SVM), artificial neural network (ANN), random forest and GBM | GBM: AUC: 0.78 | Logistic regression (LR) and least absolute shrinkage and selection operator (LASSO) regression: AUC: 0.77 |
Hanko 2021 | Undergoing primary decompressive craniectomy | 6 months | Functional outcome (GOS 1–3 vs. 4–5) | 40 (adult) | Random forest | AUC: 0.873 | Not included |
Lu 2015 | Moderate to severe TBI surviving ≥14 days | 6 months | Functional outcome (GOS 1–3 vs. 4–5) | 115 (adult) | ANN, naïve Bayes, and decision tree | ANN: AUC: 0.961 Sens: 83.5% Spec: 89.7% | LR: AUC: 0.925 Sens: 81.1% Spec: 90.1% |
Pease 2022 | GCS ≤ 8 admitted to hospital | 6 months | Functional outcome (GOS 1–3 vs. 4–5) | 757 (adult) | Convolutional neural network (CNN) | Fusion model (clinical and imaging model) AUC: 0.68 Accuracy: 82% Sens: 77.7% Spec: Varied to match each human (mean 80.3%) | International Mission for Prognosis and Analysis of Clinical Trials in TBI (IMPACT) score: AUC: 0.83 Expert humans: Accuracy: 72.7% Sens: 70.3% Spec 80.3% |
Arefan 2023 | Blunt injury with post-resuscitation GCS ≤ 8 | 24 months | Functional outcome (GOS 1–3 vs. 4–5) | 395 (with complete 24-month outcome data) | SVM, ANN, decision tree, naïve Bayes | SVM: AUC: 0.82 Accuracy: 74% | LR: AUC: 0.82 Accuracy: 71% |
Guiza 2013 | Undergoing ICP monitoring in ICU | 6 months | Functional outcome (GOS 1–3 vs. 4–5) | 160 (adult) | Gaussian processes (GP) | Core IMPACT variables with dynamic predictors: AUC: 0.87 Accuracy: 86% Sens: 88% Spec: 86% | LR: AUC: 0.68 Accuracy: 67% Sens: 67% Spec: 67% Corticoid Randomization after Significant Head Injury (CRASH): AUC: 0.67 Accuracy: 66% Sens: 67% Spec: 66% IMPACT: AUC: 0.67 Accuracy: 65% Sens: 72% Spec: 60% |
Chen 2024 | GCS ≤ 8 admitted to hospital with bloodwork drawn <12 h | 6 months | Functional outcome (GOS 1–3 vs. 4–5) | 800 (adult) | GBM, ANN, distributed random forest, generalized linear model | Distributed random forest with X1 feature (derived index with subdural hematoma thickness and coagulopathy): AUC: 0.999 | LR using CRASH variables (not exact prediction model) AUC: 0.844 LR using IMPACT variables (not exact prediction model) AUC: 0.718 |
Baucher 2019 | Acute SDH undergoing craniotomy/burr hole evacuation | 6 months | Functional outcome (GOS 1–3 vs. 4–5) | 82 (adult) | Classification and regression tree | Accuracy: 76% | Not included |
Pourahmad 2019 | Abbreviated injury severity score (AIS) score ≥3 in head region | 6 months | Functional outcome (GOS-E 1–4 vs. 5–8) | 741 (mixed; age >14 included) | SVM (linear kernel) with multiple variable selection methods | Optimal feature selection method: sequential forward selection AUC: 0.737 Accuracy: 79.1% Sens: 58.9% Spec: 89.2% | Not included |
Pourahmad 2016 | GCS ≤ 10 admitted to ICU | 6 months | Functional outcome (GOS-E 1–4 vs. 5–8) | 410 (mixed; age >11 included) | Decision tree and hybrid model using decision tree with ANN | Hybrid ANN and decision tree: AUC: 0.705 Sens: 55.1% Spec: 93.6% | Not included |
Nourelahi 2022 | Severe TBI without further specification | 6 months | Functional outcome (GOS-E 1–4 vs. 5–8) | 1682 (adult) | Random forest and SVM | SVM: AUC: 0.82 Accuracy: 78% Sens: 78% Spec: 78% | LR: AUC: 0.83 Accuracy: 78% Sens: 78% Spec: 78% |
Haveman 2019 | Moderate to severe TBI anticipated to stay in ICU > 24 h | 12 months | Functional outcome (GOS-E 1–2 vs. 3–8 and 1–4 vs. 5–8; results for latter presented) | 57 (adult) | Random forest | Electroencephalography (EEG), ICU and IMPACT variables: AUC: 0.76 Sens: 89% Spec: 80% | IMPACT score: AUC: 0.87 Sens: 100% Spec: 75% |
Tewarie 2023 | Moderate to severe TBI anticipated to stay in ICU > 24 h | 12 months | Functional outcome (GOS-E 1–3 vs. 4–8) | 104 (mixed; age >12 included) | Random forest | IMPACT and EEG features: AUC: 0.89 Sens: 83% Spec: 85% | IMPACT score: AUC: 0.81 Sens: 86% Spec: 70% |
Pang 2007 | Closed injury admitted to ICU with GCS ≤ 8 | 6 months | Functional outcome (GOS categorized as 3 levels and 5 levels; results for latter provided) | 513 (adult) | Discriminant analysis, decision tree, Bayesian network, and ANN | Discriminant analysis: Accuracy: 69% | LR: Accuracy: 62% |
Bark 2024 | TBI all severities admitted to ICU (83% model development dataset moderate to severe TBI) | 6–12 months | Functional outcome (GOS-E 1–4 vs. 5–8 and GOS-E as ordinal variable 1–8; results for latter presented) | 1808 (adult) | Random forest & ANN | Top model ANN with IMPACT features: External Dataset 1: Accuracy: 52% Sens: 52% Spec: 75% External Dataset 2: Accuracy: 21% Sens: 21% Spec: 86% | Multivariable proportional odds LR with IMPACT features: External dataset 1: Accuracy: 44% Sens: 44% Spec: 83% External dataset 2: Accuracy: 18% Sens: 18% Spec: 84% |
Eiden 2019 | Post-resuscitation GCS ≤ 8 undergoing ICP monitoring | 6 months | Functional outcome (GOS as continuous outcome) | 26 (adult) | Partial least squares regression and long-short-term memory network (LSTM) | LSTM: R2 0.732 (since GOS maintained continuous) | Not included |
Stein 2012 | Severe TBI undergoing ICP monitoring, excluding severe polytrauma | 3 months | Functional outcome (GOS-E 1–4 vs. 5–8) | 52 (adult) | Compound covariate predictor, linear discriminant analysis, k-nearest neighbor (KNN) classifiers, and SVM (differing kernels) | SVM: Accuracy: 58% | Not included |
Rubin 2019 | Severe TBI surviving >24 h (excluding patients with fixed and dilated pupils) | 6 months | Functional outcome (GOS 1–3 vs. 4–5) | 630 (adult) | Random forest, linear discriminant analysis | Random forest: AUC: 0.83 Sens: 80% Spec: 80% | LASSO regression: AUC: 0.85 Sens/spec not included for comparator model |
Farzaneh 2021 | Blunt moderate to severe TBI (excluding patients with fixed and dilated pupils) | 6 months | Functional outcome (GOS-E 1–4 vs. 5–8) | 831 (adult) | eXtreme Gradient Boosting (XGBoost) | XGBoost: Accuracy: 74.9% AUC: 0.809 Sens: 71.3% Spec: 77.5% | Not included |
Minoccheri 2022 | GCS 4–12 meeting PROTECT trial72 criteria | 6 months | Functional outcome (GOS-E 1–4 vs. 5–8) | 833 (adult) | Tropical geometry-based Fuzzy Neural Networka, XGBoost, random forest, SVM | Random forest: AUC: 0.802 Accuracy: 74.4% Sens: 57.9% Spec not included | Not included |
Folweiler 2020 | Blunt TBI meeting COBRIT trial criteria | 6 months | Functional outcome (GOS-E in unsupervised phenotype cluster analysis) | 1598 (adult) | KNN | No outcome prediction performance metric. Reported correlation of phenotype clusters to GOS-E outcomes | |
Rizoli 2016 | Blunt severe (GCS ≤ 8) TBI without hypovolemic shock surviving > 24 h | 6 months | Functional outcome (GOS-E 1–4 vs. 5–8) | 1089 (adult) | Decision tree (derived using binary recursive partitioning) | Decision tree: AUC: 0.67 Accuracy: 68.3% Sens: 72.3% Spec: 62.5% | IMPACT (extended): AUC: 0.69 Accuracy: 73.2% Sens: 92.7% Spec: 44.3% |
