Falls are the leading cause of fatal and non-fatal injuries among seniors with serious and costly consequences. Laboratory evidence supports the view that impaired ability to execute compensatory balance reactions (CBRs) or near-falls is linked to an increased risk of falling. Therefore, as an alternative to the commonly used fall risk assessment methods examining spatial-temporal parameters of gait, this study focuses on the development of machine learning-based models to detect multidirectional CBRs using wearable inertial measurement units (IMUs). Random forest models were developed based upon the data captured by five wearable IMUs to 1) detect CBRs during normal gait, and 2) identify the type of CBR (eight different classes). A perturbation treadmill (PT) was employed to systematically elicit CBRs (i.e. PT-CBRs) during walking in different directions (e.g slip-like, trip-like, and medio-lateral) and amplitudes (e.g., low-, high-amplitude). We hypothesized that these PT-CBRs could simulate naturally-occurring CBRs (N-CBRs). Proof-of-concept testing in 9 young, healthy adults demonstrated accuracies of 96.60% and 80.64% for the PT-CBR detection and type identification models, respectively. Performance of the detection model was tested against a published dataset (IMUFD) simulating N-CBRs, including the most common types observed in older adults in long-term care facilities, which achieved sensitivity of 100%, but poor specificity. Adding normal gait data from IMUFD for training improved specificity, indicating treadmill walking alone is insufficient exemplar data. Perturbation treadmill combined with overground walking data is a suitable paradigm to collect training datasets of involuntary CBR events. These findings suggest that accurate detection of naturally-occurring CBRs is feasible, and supports further investigation of implementing a wearable sensor system to track naturally-occurring CBRs as a novel means of fall risk assessment.