PANDIA: Out-of-Distribution Validation for Infant Pain Assessment

2,847 infant records across four datasets were used to validate the PANDIA multimodal AI system for pain assessment in neonates aged 0–3 months. The system, developed by Oussama El Othmani and published on figshare in May 2026, achieved 87.3% accuracy and a 92.1% clinician acceptance rate for explanations. All code, trained models, and preprocessing pipelines are publicly available on GitHub.

Use Cases

• Validate multimodal AI models for infant pain assessment based on hierarchical representation learning and graph-based inter-modal reasoning.
• Benchmark federated learning frameworks for privacy-preserving multi-site collaboration in clinical settings.
• Test meta-learning adaptation for personalized assessment with minimal per-infant data.
• Evaluate clinical interpretability of AI systems based on concept-bottleneck explanations and symbolic reasoning.
• Assess edge deployment feasibility of AI models with fewer than 30M parameters.

Strengths

• Validation involves 2,847 infants across four datasets, providing a substantial sample size.
• The system achieved a 12.4% accuracy improvement over the best baseline and performed consistently on an independent out-of-distribution test set.
• All code, trained models, and preprocessing pipelines are fully publicly available without restriction.
• The dataset is licensed under CC-BY-4.0, permitting open reuse.

Limitations

• The validation design is retrospective, and dataset heterogeneity across collection sites may introduce bias.
• Row count and column-level documentation are unknown, limiting suitability assessment.
• The primary NICU-MM dataset is available only upon request subject to an ethical data use agreement, restricting immediate access.

Provenance

Source

Oussama El Othmani via figshare.

Collection Method

Likely collected from clinical settings for research validation.

Freshness

Last updated 2026-05-26 17:32:07.