Structured & Multimodal AI for Healthcare

Clinical notes describe temporal changes, treatment responses, contradictions, diagnoses, and interventions, but sequence representations leave many of these relationships implicit. HERMES constructs patient-specific knowledge graphs from clinical narratives so that clinically meaningful relations become explicit before outcome prediction.

• · Across MIMIC-III and MIMIC-IV mortality and readmission prediction, the graph-based representations outperform the evaluated text-only baselines.
• · Clinical narratives can become more informative when important patient dynamics are represented explicitly rather than left entirely implicit in a token sequence.

Multimodal EHR models often aggregate heterogeneous observations into a compact patient representation before their task-specific contributions are known. MCPL instead learns multiple latent patient-state components from time-series and clinical-note evidence and keeps them separate until outcome-dependent aggregation.

• · Higher-weighted components retain more predictive information than random components, and removing different components changes mortality and readmission predictions by different amounts.
• · Patient state does not have to be represented by one monolithic embedding before the model determines what matters for the target outcome.

Biomedical knowledge graphs can expand observed clinical concepts into many connected candidates, but connectivity alone does not show which knowledge is supported by the current ICU record. PIKE uses temporal and note evidence to select patient-supported knowledge before integrating it with the original EHR representations.

• · The ongoing study evaluates patient-evidence-guided knowledge selection under sparse and incomplete clinical observations, with a larger advantage when observed evidence becomes sparse or the observation window becomes shorter.
• · External knowledge should be selected from current patient evidence, not treated as relevant simply because it is graph-connected.

Lumbar MRI contains complementary sequences and imaging planes. Sagittal views capture longitudinal anatomy, while axial images provide per-level geometry that is important for several findings. GateSPINE preserves sagittal and axial evidence as separate representations and combines them with content-adaptive cross-view fusion before report generation.

• · The strongest clinical-efficacy performance among the evaluated methods across three datasets, with gains largely driven by recall. The cross-view mechanism is evaluated only on cohorts containing both sagittal and axial imaging.
• · Complementary imaging views should remain distinguishable until the model can determine which evidence is useful for the current finding.

Standard 3D CT report generators typically retain one decoding path, which can create a recall bottleneck when clinically relevant pathologies are omitted. MDEF preserves multiple report hypotheses and consolidates their complementary pathology evidence before producing one final report.

• · Large recall gains over the strongest evaluated single-report baseline across in-distribution, cross-dataset, and cross-institution evaluation, distributed across many pathology classes — though the gain is recall-driven and comes with a precision trade-off.
• · Multiple hypotheses can improve clinical coverage, but completeness must be balanced against precision.