Machine learning models map treatment decisions in posterior communicating artery aneurysms
Posterior communicating artery aneurysms sit at a difficult intersection of anatomy, rupture risk, and procedural trade-offs. In a 75-patient institutional cohort treated between 2002 and 2023, investigators built and externally validated machine learning models to predict treatment selection, complications, intraoperative rupture, angiographic obliteration, and discharge functional status. The project focused on a specific clinical problem: whether preoperative variables already available in routine practice can help forecast which patients are steered toward open surgery or endovascular therapy, and how those choices relate to short-term outcomes.
The model set was broad rather than narrow. Five algorithms were compared — Random Forest, Support Vector Machine, Logistic Regression, Decision Tree, and XGBoost — across seven outcomes: treatment type, specific open surgical treatment, specific endovascular treatment, intraoperative rupture, complications, angiographic obliteration, and discharge modified Rankin Scale score. Preoperative inputs included age, sex, family history, aneurysm size, presentation, morphology, and location. Before modeling, the data underwent class balancing, feature engineering, and regularization to reduce overfitting, a necessary step in a dataset where rare events can otherwise vanish into the majority class.
Treatment selection was forecast with useful but imperfect accuracy
For treatment prediction, the best-performing model reached 78.7% accuracy with Random Forest. In practice, that level of performance suggests the models captured substantial decision structure embedded in the clinical record, but not enough to replace specialty judgment at the bedside. Open versus endovascular selection in posterior communicating artery aneurysms reflects both lesion geometry and the operator’s estimation of procedural risk, so a modest-to-strong classifier should be read as decision support rather than automation.
A single accuracy number can obscure the more important issue of which events a model actually sees. For complications, the Random Forest model produced 94.7% accuracy, yet sensitivity was 0%, making the score clinically misleading for event detection. A model that misses every complication is not useful for safety surveillance, no matter how polished the overall percentage appears. In aneurysm care, where the absolute number of adverse events is often small, class imbalance can create the illusion of performance unless recall is examined alongside accuracy.
Rare events improved only after class balancing
The sensitivity analyses are the most instructive part of the study. When class balancing was applied, minority class recall improved from 0.0% to 52.2%–86.3% for rare outcomes. In plain terms, the algorithms began to recognize events they had previously ignored. That shift matters because aneurysm rupture, procedural complications, and poor functional recovery are exactly the outcomes that most influence patient counseling and treatment planning, yet they are also the outcomes most likely to be underrepresented in training data.
Cross-validation performance was particularly strong for intraoperative rupture and angiographic obliteration. Random Forest achieved 92.0% accuracy for intraoperative rupture prediction and 87.3% accuracy for angiographic obliteration. Those outputs fit the clinical intuition that aneurysm size, morphology, and presentation often track with technical difficulty and the likelihood of complete exclusion from the circulation. Even so, performance metrics alone do not reveal whether the model is learning robust biology and procedure-related risk, or simply pattern-matching a single-center practice style.
External validation exposed the limits of local prediction
External validation was the real test. An independent cohort from the University of Louisville, with 20 most complete cases for the primary external cohort and up to 28 cases for selected endpoints with available labels, showed variable generalizability. Treatment prediction remained strong at 78.1% accuracy, but complication prediction fell to 70.3% externally. In a smaller, separate patient population, that drop is a reminder that institutional technique, case mix, perioperative pathways, and documentation practices all shape predictive behavior.
Neurovascular machine learning models often look strongest inside the dataset that created them. Once they cross into another center, the signal can narrow because aneurysm management is not only a matter of anatomy; it is also a matter of surgeon preference, device availability, patient referral patterns, and local thresholds for intervention. The external validation here makes the study more credible, not less, because it shows where the models travel well and where they do not.
Age, aneurysm size, and neurologic grade carried the most weight
Feature importance analysis repeatedly elevated age, aneurysm size, admission modified Rankin Scale score, and Hunt-Hess grade. Those variables are unsurprising to any clinician who manages intracranial aneurysms, but their repeated prominence is still useful. Age and aneurysm size may reflect cumulative vascular fragility and the technical consequences of larger lesions, while admission modified Rankin Scale score and Hunt-Hess grade capture the patient’s neurologic reserve and the severity of presentation. In a disease where a small change in presentation can alter both the choice of repair and the expected recovery curve, these four inputs carry practical weight.
The biological logic is straightforward. Larger aneurysms and worse neurologic grades tend to mark either more unstable lesions or more advanced hemorrhagic injury, and those conditions often push management toward different procedural strategies and different expectations for obliteration and recovery. The model does not replace that logic; it formalizes it into a quantitative framework that can be tested, audited, and refined.
Why this framework matters for bedside neurovascular care
Posterior communicating artery aneurysms account for roughly 25% of intracranial aneurysms, and they remain a difficult subgroup because treatment selection is rarely binary. An open operation may offer durable exclusion but at the price of greater surgical exposure, while endovascular therapy may be less invasive but may not fit every neck configuration, morphology, or rupture scenario. A validated prediction tool that integrates patient and aneurysm features could help teams compare expected complication risk, likelihood of obliteration, and functional outcome before the procedure begins.
Still, the study’s own data define the ceiling of immediate clinical use. The cohort was small, the work came from an institutional database, and external validation, while important, was limited in size. Complication prediction in particular showed how easily a high accuracy score can mask poor event detection if sensitivity is not strong. For now, the most defensible use case is decision support in combination with expert neurovascular judgment, not standalone triage.
The study represents the first externally validated, multioutcome machine learning framework for posterior communicating artery aneurysms, but its clinical value will ultimately depend on replication in larger and more diverse cohorts, with prospective testing against real-world treatment decisions and outcomes.