Introduction: Predicting vestibular schwannoma (VS) tumor growth (>=2mm) could facilitate risk stratification to tailor management strategies, mitigating unnecessary long-term monitoring in low-risk patients. Existing models rely primarily on baseline features, which fail to capture heterogeneous disease trajectories and lack time-to-event analysis. We aimed to develop a machine learning-based model that utilizes diverse longitudinal radiographic and clinical features from patient records to dynamically predict risk of tumor growth in untreated VS patients.
Methods: A retrospective review of health records from a single neurosurgical center identified adult patients with VS between January 2000 and August 2024. Inclusion required management via the W&S approach. Radiographic characteristics (size, enhancement, texture) and clinical symptoms (hearing impairment, cranial nerve dysfunction, audiology results) were extracted as dynamic features from all available longitudinal MRI reports and clinical notes describing VS from baseline until active treatment initiation/end of follow-up. Static demographic features were recorded at baseline. Using the ‘dynamicLM’ tool in R, we constructed a Cox landmark model to predict risk of tumor growth of >=2mm within a specific duration from baseline and each subsequent risk assessment 'landmark' time. Prediction accuracy was evaluated using time-varying area under the curve (AUC) and calibration across a series of landmarks.
Results: Overall, 306 patients were identified. Mean age was 59.9 years (standard deviation [SD]=14.7), there was a higher proportion of women (N=164,53.6%) and median follow-up duration was 43.2 months. Mean maximum tumor dimension was 16.7mm (SD=9.3), 44.8% of tumors involved the cerebellopontine angle, 22.9% caused brainstem compression, 21.9% were heterogeneous, 77.1% were enhancing and 4.2% had associated edema. 88% of patients presented with hearing loss, 51.0% with tinnitus, 20.0% with facial symptom/signs and 45.0% with vestibular/cerebellar dysfunction. The final model achieved good calibration (slope of 0.9-1.2) and high discrimination (AUC>0.8) across landmarks.
Conclusion : We developed a ML model to accurately predict the dynamic risk of tumor growth in VS within specific time frames using a diverse set of longitudinal features. This model may enhance personalized risk predictions to aid decision-making when managing VS patients.
.jpg)
