What the programme covers
Multimodal data is valuable only when the components describe the correct patient and clinical episode. Joining records too broadly can introduce label leakage, future information or false relationships between examinations and outcomes.
What a useful specification includes
A defensible request defines the clinical task, source evidence and acceptance criteria before patient-level data moves. The exact fields and thresholds depend on the intended model claim.
- Modalities and clinical fields required
- Patient, encounter and temporal linkage rules
- Index event and prediction horizon
- Missing-modality and partial-record policy
- Label, outcome and censoring definitions
- Training and validation separation
Quality and validation controls
The programme data graph records which sources were joined, which were missing and what transformations produced each model input. This makes temporal and cross-modal leakage inspectable.
- Identity-safe cross-system resolution
- Time-window and episode consistency
- Missing-modality distributions
- Duplicate and conflicting evidence
- Feature, label and future-information leakage
Availability, rights and delivery
Multimodal releases can be represented as linked manifests, files and structured tables with pseudonymous keys, data dictionaries and transformation lineage.
Public pages describe a sourcing and engineering capability, not guaranteed ready inventory. Each release remains subject to verified programme inventory, programme-specific authorization, privacy review, technical acceptance and buyer licence terms.
Questions, answered directly.
Does multimodal mean every case has every data type?+
No. Completeness should be reported, and the model design must define how missing modalities are handled.
Can images and reports be paired?+
Yes after verifying the report belongs to the correct examination and final version.
Can multimodal data train foundation models?+
Potentially, subject to programme authority, technical feasibility and the intended use.
How is label leakage controlled?+
By defining the prediction time, limiting future information and reviewing derived features against the target outcome.