Advances in Machine Learning Paradigms for Reliable Swimming Pool Drowning Detection: Systematic Review of Methods and Implementations

Drowning is a leading cause of unintentional injury-related mortality worldwide, particularly in swimming pool environments, where rapid detection is critical to prevent fatalities. Traditional monitoring approaches rely on lifeguards and rule-based systems, which are often limited by human error, delayed response, and environmental challenges such as occlusion, glare, and multiple swimmers. Machine learning (ML) and computer vision methods have emerged as promising alternatives for automated drowning detection; however, their methodological rigor, real-world applicability, and reliability remain unclear.

This study aimed to systematically review advancements in ML-based drowning detection systems over the past 15 years. The review focused on detection methods, sensing modalities, evaluation metrics, and validation practices. It also examined how studies addressed real-world challenges such as occlusion, lighting variability, and multiswimmer environments, and identified key gaps related to generalizability, reliability, and privacy.

A systematic literature review was conducted using major electronic databases, including IEEE Xplore, Scopus, PubMed, Web of Science, ACM Digital Library, and Google Scholar. A total of 4927 records were initially identified and screened by 2 independent reviewers using predefined inclusion and exclusion criteria. A final set of 30 studies was included for data extraction and synthesis. The methodological quality and risk of bias were assessed using the Prediction Model Risk of Bias Assessment Tool (PROBAST), while reporting transparency was evaluated using the TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) guidelines. Data analysis was conducted in R.

The findings indicate that drowning detection research is dominated by supervised, vision-based approaches (23/30, 76.7%), with convolutional neural networks (5/30, 16.7%) and You Only Look Once-based architectures (7/30, 23.3%) being the most commonly used methods. While reported performance is generally high, with mean accuracy exceeding 90.9%, evaluation practices are heavily centered on accuracy, with limited reporting of precision, recall, false alarm rate, and latency. One of 30 (3.3%) studies reported validation under real-world operational conditions, with an additional 5 (16.7%) studies using real-world data without formal real-world validation, and no studies conducted cross-dataset evaluation. In addition, critical real-world challenges such as occlusion, glare, and crowding were rarely addressed (<10% of studies), and all studies relied on private datasets. The PROBAST assessment further indicated a high risk of bias in 96.7% (29/30) of studies, primarily due to limited validation and nonrepresentative data sources.

ML-based drowning detection systems have demonstrated substantial progress in algorithm development, particularly in vision-based models. However, this progress is not matched by equivalent advances in validation, robustness, and real-world applicability. The field is characterized by a significant gap between experimental performance and deployment readiness. Future research should prioritize standardized evaluation frameworks, external validation, comprehensive performance reporting, and the integration of multimodal and privacy-preserving approaches to enable reliable real-world implementation.