<p>Hearing loss is an impairment that affects a large population worldwide. Cochlear implant (CI) surgery is an effective solution, especially in patients with residual, natural hearing. However, the risk of intracochlear damage during CI surgery, which impairs residual hearing, remains as a concern. It has been shown that intra-operative monitoring (using electrocochleography, specifically the cochlear microphonic) during CI surgery can detect drops that are associated with such trauma, resulting in its potential prevention. Researchers have attempted to automate the process of drop detection by implementing supervised models. However, these have so far failed to achieve good reliability and generalizability. To address these issues, we develop an unsupervised model based on the Bayesian method with a scoring function derived from the logistic function. The proposed model, as a data-driven method, directly learns from data without the need for any supervision which enhances its adaptability to diverse patient populations. Empirical results demonstrate the proposed model’s high accuracy with a low false alarm rate (18 extra detection against 146 in previous work), as well as its robustness when processing new data. This indicates its potential for real-world application to improve safety during CI surgery.</p>