Temporal Expression Normalization

The core computational process of resolving relative temporal expressions by anchoring them to a known reference point. When a clinical note states 'three weeks post-discharge,' the system identifies the discharge date as the anchor and calculates the absolute calendar date. This requires document creation time (DCT) awareness and the ability to traverse temporal relations across multiple sentences. The process handles expressions like 'tomorrow,' 'last month,' and 'the following Tuesday' by establishing a temporal focus and applying calendar arithmetic. Failure in anchoring leads to cascading errors in medication schedules and follow-up reminders.

The systematic mapping of clinical frequency abbreviations to standardized ISO 8601 duration formats. Common expressions include:

• q.d. → R/P1D (once daily)
• BID → R/P12H (twice daily)
• TID → R/P8H (three times daily)
• q.i.d. → R/P6H (four times daily)
• q.o.d. → R/P2D (every other day)
• qHS → nightly at bedtime

This conversion enables automated scheduling systems to calculate precise administration times and detect frequency conflicts in polypharmacy scenarios.

A widely-adopted open-source temporal tagger that employs a cascade of hand-crafted rules and regular expressions to identify and normalize temporal expressions. HeidelTime operates in three phases: tokenization and part-of-speech tagging, temporal expression detection using pattern matching, and value normalization to the TIMEX3 annotation standard. It includes domain-specific resources for clinical text, recognizing expressions like 'post-op day 3' and 'gestational week 28.' The system outputs structured annotations with attributes for type (DATE, TIME, DURATION, SET), value, and modifier.

A deterministic rule-based temporal tagger built on the Stanford CoreNLP pipeline that excels at recognizing complex temporal patterns. SUTime uses finite-state transducers over regular expressions defined in TokensRegex to identify expressions like 'the third Thursday of next month' or 'biweekly on Mondays.' It integrates with semantic role labeling to resolve temporal relations between events and times. The system outputs TIMEX3 annotations and is particularly effective at handling recurring temporal sets and holiday-based expressions that appear in clinical scheduling contexts.

The ISO-TimeML markup language for representing temporal expressions in structured form. A TIMEX3 tag encodes:

• type: DATE, TIME, DURATION, or SET
• value: Normalized ISO 8601 representation
• mod: BEFORE, AFTER, APPROX, START, END, MID
• quant: EVERY, BIWEEKLY, etc.
• freq: Frequency within a period

Example: 'the past three days' becomes <TIMEX3 type='DURATION' value='P3D' mod='BEFORE'/>. This standard enables interoperability between different temporal normalization systems and provides a consistent format for downstream clinical reasoning engines.

The process of identifying TLINKs (temporal links) between events and times in clinical narratives. After normalization, systems establish relations like BEFORE, AFTER, OVERLAP, and CONTAINS to build a temporal graph of the patient's clinical timeline. For example, 'The patient developed a fever after receiving the vaccine' creates a TLINK where the fever event occurs AFTER the vaccination event. This graph enables chronological ordering of clinical events, critical for understanding disease progression and treatment response across multiple encounters.