Back to 2025 Abstracts
Anderson Mandibular Osteoradionecrosis Speech And Swallow Tool (AMOSST): A Validated, Highly Predictive, Machine Learning-powered Predictor Of Postoperative Outcomes
Rami Elmorsi, MD, Z-Hye Lee, MD, Lucia Martinez-Sanchez, BS, Hrishabh R. Bhosale, BA, Austin Ha, MD, Patrick B. Garvey, MD, Edward I. Chang, MD, Rene D. Largo, MD, Matthew M. Hanasono, MD, Peirong Yu, MD.
MD Anderson Cancer Centre, Houston, TX, USA.
Introduction: Mandibular osteoradionecrosis (ORN), a severe complication of head and neck radiation, often mandates extensive locoregional resections and reconstructions. Accurate prediction of postoperative speech and swallow function is essential for patient counseling and postoperative care.
Methods: A 30-year retrospective study of mandibular ORN reconstructions from 1994 to 2023 established the largest database of its kind, which was in-turn leveraged to construct AMOSST. Of 278 ORNs, 229 (85%) were used to tune, train, and validate three machine learning models (Lasso Regression, Gaussian Bayes, and FasterRisk) with leave-one-out and 10-fold cross-validation to prevent overfitting. The remaining 49 (15%) ORNs were reserved for external validation to ensure generalizability.
Results: Two Lasso regression models excelled in predicting intelligible speech and return to oral feeds postoperatively, achieving 92% and 80% accuracy, respectively, on external validation (Table 1). Both models were subsequently transformed into nomographic calculators to facilitate broader clinical application and utility (Figure 1, Speech; Figure 2, Swallow).
Conclusion: AMOSST is a validated, powerful predictive tool that is developed from the largest mandibular ORN database to date.
Performance of Anderson Mandibular Osteoradionecrosis Speech and Swallow Tool (AMOSST) | Area under ROC Curve | Classification Accuracy | F1-Score | Precision | Recall |
Predicting >50% intelligible Speech |
10-fold cross-validation | 0.95 | 0.89 | 0.89 | 0.89 | 0.89 |
Leave-one-out cross validation | 0.95 | 0.88 | 0.88 | 0.88 | 0.88 |
External validation | 0.98 | 0.92 | 0.93 | 0.94 | 0.92 |
Predicting Tolerance of Oral Feeds |
10-fold cross-validation | 0.82 | 0.74 | 0.74 | 0.74 | 0.74 |
Leave-one-out cross validation | 0.83 | 0.76 | 0.75 | 0.76 | 0.76 |
External validation | 0.76 | 0.80 | 0.80 | 0.80 | 0.80 |
Back to 2025 Abstracts