Table 2.
Summary of findings for existing studies applying machine learning to predict unruptured intracranial aneurysm ‘stability’. The authors defined stability as a composite outcome which included rupture as well as aneurysm growth and/or presence of symptoms
| Publication | Study Design | Modality | Reference standard | Time period of risk assessment | Definition of stability | Comparison to clinical practice | Index test | Model features | Hold-out test set (n) (or other specified dataset) |
Hold-out test set performance accuracy (or performance of other specified dataset) |
|---|---|---|---|---|---|---|---|---|---|---|
|
Liu et al. 2019 [28] |
Development and prospective validation | DSA | Comparison to result (stable/unstable) | 1 month stability assessment (follow-up median: 11.5 months, range: 3–26 months) |
1. Remained unruptured 2. No UIA growth 3. Asymptomatic |
– |
1. Generalized linear model* 2. Ridge regression 3. Logistic regression |
(1) Clinical features (2) Morphological features |
IV: 124 | AUC: 0.86 |
|
Zhu, et al. 2020 [29] |
Development and retrospective validation | 3D-DSA | Comparison to result (stable/unstable) | 1 month stability assessment (median follow up 15.6 months; range 5–39 months) |
1. Remained unruptured 2. No UIA growth |
– |
1. Neural Network* 2. Random forest 3. SVM |
(1) Clinical features (2) Morphological features |
IV: 411 |
Accuracy: 0.82 Balanced Accuracy: 0.72 Sensitivity: 0.52 Specificity: 0.93 AUC:0.87 PPV: 0.71 NPV: 0.85 F1 Score: 0.60 |
|
Yang, et al. 2021 [30] |
Development and prospective validation | CTA | Comparison to result (stable/unstable) | 3 years |
1. Remained unruptured 2. UIA Growth ≤ 20% |
Comparison made to PHASES, ELAPSS, UIATS and IARS Score | 1. Neural network |
(1) Clinical features (2) Morphological features (3) Hemodynamic features |
IV: 37 (9-fold cross validation on training data, no hold-out test set per se) |
AUC: 0.83 |
| Liu, et al. 2022 [31] | Development and prospective validation | CTA | Comparison to result (stable/unstable) | 2 years |
1. Remained unruptured 2. UIA of aneurysm < 20% or < 1 mm |
Comparison made to PHASES and ELAPSS | 1. Logistic regression |
(1) Clinical features (2) Morphological features (3) Hemodynamic features |
IV: 97 | AUC: 0.94 |
| Zhang, et al. 2023 [32] | Development and retrospective validation |
CTA/ MRA |
Comparison to result (stable/unstable) | 2 years |
1. Remained unruptured 2. No IA growth 3. Asymptomatic |
– |
1. SVM* 2. Logistic regression 3. Adaboost |
(1) Hemodynamic features | EV: 54 |
Accuracy: 0.83 Balanced Accuracy: 0.83 Sensitivity: 0.83 Specificity: 0.83 AUC: 0.89 PPV: 0.71 NPV: 0.91 F1 Score: 0.77 |
| Irfan, et al. 2023 [33] | Development only | DSA | Comparison to risk interpretation of the UIAs was conducted by expert neurosurgeons | Duration not stated | Not defined | Comparison to neurosurgeon expert opinions for reference standard |
Combination model: 1. Neural network 2. Decision tree classifier |
(1) Morphological features (2) Hemodynamic features |
IV: 141 |
Accuracy: 0.85 Balanced Accuracy: 0.85 Sensitivity: 0.84 Specificity: 0.86 AUC: 0.93 PPV: 0.82 NPV: 0.87 F1 Score: 0.83 |