Objectives: Multiple factors and criteria are considered by health technology assessment (HTA) agencies, resulting in discordant recommendations; limited analyses exist on what drives differences in HTA recommendations for Medical Device technologies. Using machine learning, we explored 1) if available factors can predict the appraisal decision and 2) if factors that contribute most to differences in appraisal decisions can be identified. Methods: Medical device HTA appraisals with a decision date between 01 December 2003 and 08 October 2019 from the IQVIA^TM HTA accelerator were analyzed. Natural language processing was first used to extract keywords as features from data elements for clinical outcomes and conclusions. A machine learning algorithm, XGBoost, was then used to predict positive or negative HTA recommendations. Ten-fold cross-validation was used for training models and testing performance. To make the data more balanced, synthetic minority over-sampling was used to adjust the ratio between positive and negative recommendations. Results: 1,430 HTA submissions were included, representing 25 therapeutic areas, 10 HTA agencies, and 9 countries globally. The top 5 predictors were: primary therapeutic area, HTA agency, level of evidence (e.g. systematic review being the most important and expert opinion the least important), HTA type (procedure/intervention or medical device assessments), and the appearance of key string: “statistically significant difference”. Prediction performance, based on area under the receiver operating characteristic (ROC) curve, was 0.80. Conclusion: Machine learning could help predict HTA appraisal decisions and provides direct insight into relative contributions to prediction accuracy. Models of this nature may be useful for early screening of medical devices and interventions to predict device recommendations. Though therapeutic areas and HTA agency were important features of the model, explanatory research is required to better understand interactions among these and other model inputs.