Temporal Concept Grounding

The process of identifying the chronological link between a clinical event and a temporal expression. This involves classifying the relationship using standards like THYME or TLink tags.

• BEFORE: The condition occurred prior to a reference point (e.g., 'history of MI').
• AFTER: The condition occurred subsequent to a reference point.
• OVERLAP: The event occurs during the specified interval.
• BEGINS/ENDS: The event marks the start or termination of a period.

The conversion of relative or vague clinical expressions into a machine-readable, standardized format. This step is critical for computational reasoning.

• Absolute Dates: '2023-10-26' for a specific encounter.
• Durations: 'P6M' for a six-month treatment course.
• Fuzzy Intervals: Mapping 'a few weeks ago' to a probabilistic range like '2023-10-01/2023-10-14'.
• Admission/Discharge Anchors: Using the visit timestamp as a reference point for 'on admission'.

The downstream task of assigning a clinical status to a grounded entity based on its temporal context. This moves beyond simple presence/absence to a nuanced patient history.

• Active: The problem is currently affecting the patient.
• Resolved: The condition is no longer present.
• History Of: The patient experienced this in the past but it is not active.
• Recurrent: The condition is episodic and has returned.
• Chronic: A long-lasting condition that is currently managed.

The architectural pattern of aggregating temporally grounded entities from multiple clinical notes into a single, unified patient timeline. This requires resolving co-references and conflicting dates.

• Longitudinal Record: A single view of a patient's disease trajectory over years.
• Conflict Resolution: Heuristics or models to decide the truth when two notes disagree on a surgery date.
• Temporal Scoping: Limiting a diagnosis to a specific encounter (e.g., 'pneumonia during hospitalization X').

Rule-based and neural methods for determining the scope of a temporal modifier. The system must decide which clinical entities a temporal expression governs.

• Sentence-Level Scoping: Assuming a 'date of onset' applies to all findings in the same sentence.
• Section-Level Scoping: Linking all entities in the 'Past Medical History' section to a 'historical' status.
• Dependency Parsing: Using syntactic trees to precisely attach a time adverb to a specific verb or noun phrase.

Leveraging a structured ontology to infer temporal constraints that are not explicitly stated. This uses logical axioms to deduce timelines.

• Transitivity: If A happened before B, and B before C, the system infers A happened before C.
• Temporal Subsumption: Knowing that a 'cholecystectomy' implies a prior diagnosis of 'gallbladder disease'.
• Constraint Propagation: Using a known medication start date to infer the minimum duration of a related condition.

The ability of transformer-based models like ClinicalBERT or GatorTron to encode the sequential nature of events. This goes beyond bag-of-words by using positional embeddings and attention masks to understand narrative flow.

• Pre-training Objectives: Models trained on temporally ordered clinical sequences learn to predict the next event.
• Temporal Probing: Diagnostic classifiers test if a model's hidden states encode the chronology of a patient's history.
• Limitation: Standard self-attention is permutation-invariant; explicit temporal position encoding is often required.