Grounded in evidence.
Built for the real world.

ML models consistently achieve AUC >0.80 for high-acuity ED outcomes including ICU transfer and hospital admission (Rajpurkar et al., 2022; Topol, 2019).

NLP on free-text triage notes significantly improves classification accuracy over structured data alone — the core motivation for ClinAssist's dual-model architecture.

XGBoost and gradient boosting outperform simpler models across diverse ED populations in recent systematic reviews.

Clinician trust remains the primary barrier to adoption — not model performance — underscoring why explainability is central, not optional (Lundberg & Lee, 2017).

Co-attention transformers for multimodal clinical data (Chen et al., ICCV 2022) provide theoretical grounding for ClinAssist's unified inference layer.

Attention-based NLP (Vaswani et al., 2017) forms the backbone of ClinAssist's clinical notes module, enabling both prediction and explanation from free text.

MIMIC-IV — Beth Israel Deaconess Medical Centre. 300,000+ ICU and ED admissions. Gold standard for clinical AI model development and validation.

PhysioNet — Open repository of physiological and clinical data. Used for model development and cross-dataset validation alongside MIMIC-IV.

TGA guidance (Australia) — ClinAssist is designed from the ground up to align with the Therapeutic Goods Administration's AI-as-a-Medical-Device framework. Regulatory scoping is a Phase 1 deliverable.

Australian Privacy Act compliance — Data preprocessing pipelines are designed around Australian Privacy Act principles, including de-identification and consent frameworks.

IRB/Ethics approval — Formal ethics clearance for dataset access and any clinician interview or study components will be obtained in Phase 1 prior to data collection.

Clinician-first design: Every architectural decision starts with a single question — will a registrar at 3am be able to act on this? We embed clinician feedback into the design loop from day one, not after the model is built.

Multimodal from the ground up: Rather than bolting imaging or NLP onto a structured-data model, ClinAssist is architected to fuse all three inputs at the inference layer — producing a single, explainable risk output that reflects the full clinical picture.

Validation that measures trust, not just accuracy: AUC alone doesn't predict adoption. Our validation framework measures whether clinicians trust and act on ClinAssist's outputs — using human-factors methods alongside standard ML metrics.

Regulatory alignment from day one: TGA AI-as-a-Medical-Device scoping is a Phase 1 deliverable, not an afterthought. Australian Privacy Act compliance and IRB ethics approval are built into the data pipeline design, not retrofitted.