<p>Simulation-based training is widely used in neonatology to prepare healthcare professionals for high-risk clinical situations. In this context, effective communication and teamwork are critical for patient safety. However, evaluating communication quality during simulation sessions often relies on time-consuming manual review of audio and video recordings by expert trainers. Moreover, the poor acoustic quality of recordings and the presence of overlapping speech and background noise make conventional speech-recognition-based analyses unreliable. To address these limitations, this paper proposes an automated workflow that detects acoustic-prosodic anomalies associated with potentially ineffective communication in neonatal simulation sessions using syllabic-scale acoustic analysis and unsupervised deep learning. The approach extracts super-segmental features inspired by psychoacoustic principles and models them through deep neural architectures to identify anomalous dialogue segments associated with stress, uncertainty, elevated vocal effort, or inefficient interaction patterns. Unlike traditional methods relying on speech transcription or phonetic-level processing, the proposed system operates directly on syllabic-scale acoustic patterns, making it more robust to noisy and heterogeneous recording conditions. The workflow was evaluated across multiple neonatal simulation case studies and compared with a traditional cluster-analysis baseline. The proposed method achieved an average accuracy of 87.7%, with a maximum of 91.2%, representing improvements of up to 59% over the baseline approach. The system also demonstrated robustness to high noise levels and produced training-quality assessments showing strong agreement with expert evaluations. By automatically identifying acoustic-prosodic indicators associated with potentially ineffective communication, the proposed approach can support objective review and feedback processes in neonatal simulation-based training.</p>