A Benchmark Study for Reporting Feasibility in AI-Based Infant Hearing Screening: Exploring the Limits of Passive Sensing

Assessing hearing in infants with a developmental age of 3–7 months remains a clinical challenge. Too mature for neonatal reflexes, yet too immature for visual reinforcement, this demographic relies on Behavioural Observation Audiometry (BOA), a subjective method with high inter-observer variability. This preliminary study investigated whether State-of-the-Art (SOTA) machine learning could automate a modified BOA protocol, distinguishing auditory responses in a passive listening task. We analysed facial video recordings from 46 infants (mean age 4.68 months; central 95% range: 3.68–6.44). The protocol used a central “attention-holding” visual stimulus and loudspeakers at ±45°. Using OpenFace 2.0, we extracted 38 facial, gaze, and pose channels over 12 s windows (−2 s pre-stimulus to +10 s post-stimulus). We benchmarked eight time-series architectures—ranging from Logistic Regression to the MOMENT foundation model—to classify “Stimulus” (80.25%) vs. “Control/Silence” (19.75%). Despite subject-independent validation and augmentation, no model exceeded chance-level performance (Balanced Accuracy 0.50–0.60). UMAP projections revealed a complete topological overlap between stimulus and control frames, both in raw feature space and learned embedding space. Unlike our previous success with older infants (7–24 months), these findings suggest automated facial analysis during this specific passive listening paradigm may be unlikely to serve as an effective screening method for this demographic. We attribute this to insufficient stimulus salience and high inter-subject variability—both in response timing and idiosyncratic baseline movements—which obscure the transient response signal. Future work must prioritise protocol adaptations to enhance response clarity, supporting a reframing from supervised classification toward anomaly detection and clinician-validated models.