Temporality Classification

The core mechanism that identifies the chronological link between a social risk mention and the document's reference time. This involves parsing linguistic cues like verb tense, temporal adverbs, and date expressions.

• Current: 'Patient is currently homeless and living in a shelter'
• Historical: 'Patient reports a history of homelessness 2 years ago'
• Future: 'At risk of eviction next month if unable to pay rent'

This classification directly informs care team urgency and intervention triage.

Temporality classification requires analyzing a contextual window around the SDOH mention, not just the entity itself. The model must attend to surrounding sentences to resolve temporal ambiguity.

• Negation + Temporality: 'Patient denies current food insecurity' vs. 'Patient reports past food insecurity'
• Experiencer + Temporality: 'Patient's brother was homeless last year' (historical, non-patient)
• Hypotheticals: 'If patient becomes homeless, refer to social work' (conditional future)

This prevents misclassification of non-patient or hypothetical events as active risks.

Extracted temporal expressions are normalized to standardized formats for interoperability and downstream analytics. This maps relative expressions to absolute or interval-based representations.

• Relative: '3 months ago' → normalized to a specific date range
• Duration: 'homeless for 6 months' → encoded as a duration value
• Recurrence: 'intermittent housing instability' → flagged as episodic pattern

Normalization enables population-level temporal queries across patient cohorts.

The model anchors all temporal classifications to the document creation timestamp as the reference point. This is critical because clinical notes may be written days after an encounter.

• A note written on 2024-06-15 mentioning 'homeless since last Tuesday' requires resolving 'last Tuesday' to 2024-06-11
• Admission notes, progress notes, and discharge summaries each have distinct reference times
• Longitudinal analysis tracks temporality changes across multiple encounters

This anchoring prevents temporal drift in population health dashboards.

Temporality classifiers output a confidence score for each temporal label, enabling human-in-the-loop review workflows for ambiguous cases.

• High confidence: Explicit temporal markers ('currently', 'in 2019')
• Low confidence: Implicit or vague references ('has had issues with housing')
• Uncertainty flags: Triggered when temporal signals conflict within the context window

Low-confidence predictions are routed to clinical reviewers, ensuring data quality for downstream risk stratification models.

Beyond single-document classification, temporality models enable longitudinal tracking of SDOH status across a patient's entire record. This reveals patterns of chronic, resolved, or escalating social risk.

• Chronic: Consistent 'current' housing instability across multiple encounters
• Resolved: Transition from 'current' to 'historical' status after intervention
• Escalating: Progression from 'at risk' to 'current' homelessness

This longitudinal view supports closed-loop referral tracking and outcomes measurement for value-based care contracts.

A contextual NLP task that identifies who is experiencing the social risk mentioned in a clinical note. This is a critical prerequisite to temporality classification, as the system must first distinguish whether the patient or a family member/caregiver is the subject before determining the chronological status of the event.

• Distinguishes patient from proxy experiencers
• Resolves ambiguous pronoun references
• Prevents misattribution of historical risks

A contextual analysis technique that distinguishes whether a social risk factor is present or absent in clinical text. Temporality classification interacts directly with negation—a historical mention that is also negated ('patient denies past homelessness') carries fundamentally different clinical weight than an affirmed current crisis.

• Detects negation cues: 'denies', 'no evidence of'
• Scopes negation to specific spans
• Prevents false-positive risk flags

The process of using algorithms to identify patients with specific social need profiles based on a combination of structured codes and unstructured clinical data. Temporality classification enriches phenotyping by adding a chronological dimension, enabling distinctions between active, resolved, and emerging risk cohorts.

• Combines ICD-10 Z-codes with NLP outputs
• Enables temporal cohort stratification
• Supports longitudinal risk tracking

A domain-specific language model fine-tuned on clinical notes to generate contextual embeddings that improve the accuracy of social determinant extraction tasks. Fine-tuned BERT models excel at temporality classification by learning subtle linguistic cues—such as verb tense and temporal adverbs—that signal whether a risk is current, historical, or future.

• Captures tense and temporal modifier signals
• Outperforms rule-based temporal taggers
• Requires annotated temporality corpora

A semantic network that connects social risk concepts, ICD-10 codes, and community resources to enable complex reasoning over patient social data. Temporality classification feeds the graph with time-stamped edges, allowing queries like 'patients with resolved housing insecurity in the last 6 months' for outcome tracking.

• Models temporal relationships as graph edges
• Enables time-bound cohort queries
• Links historical risks to current outcomes

A detailed instruction manual for human annotators that defines the scope, entity types, and edge cases for labeling social determinants of health in a gold standard corpus. For temporality, guidelines must specify exactly how to classify mentions using a consistent framework—typically Current, Historical, Future, or Unknown.

• Defines temporal label taxonomy
• Addresses ambiguous boundary cases
• Essential for inter-annotator agreement