Disambiguating Past and Present States in Clinical Text

45 concepts31 papers1 proposal

When a medical natural language processing (NLP) model reads a patient's chart, it encounters a dense web of medical terminology. If the chart reads, "Patient has a history of depression but presents today with acute chest pain," a naive model might flag both "depression" and "chest pain" as active clinical concerns.

This is the core challenge of temporal state disambiguation: distinguishing between historical mentions, future plans, and current active clinical states within unstructured text. Because clinical narratives use identical terminology whether describing a condition from ten years ago or an emergency happening right now, models frequently suffer from severe false-positive rates, misattributing historical context as an immediate clinical risk.

The Linguistic Trap of Keyword Overlap

Why do standard text classification models struggle so much with medical timelines? The root of the problem lies in linguistic similarity.

As Yang Cui et al. demonstrate, medical documents describe historical treatments using language that is nearly identical to descriptions of current hospitalizations. For example, a list of medications a patient took five years ago looks exactly like a list of medications prescribed today. Keyword-based classification systems and automated coding pipelines frequently fail because they cannot interpret this temporal context; they see the keyword and assume the condition or treatment is currently active.

This failure mode spans multiple clinical domains:

  • Automated ICD Coding: In automated medical coding, Danchen Zhang et al. show that classifiers routinely misassign diagnostic codes because they fail to separate active diagnoses from historical mentions or speculative language used to rule out a disease.
  • Oncology Trajectories: Omid Rohanian et al. found that binary classification models struggle to differentiate between a patient's past cancer history and their current active diagnoses, which corrupts the mapping of patient trajectories over time (Kasper Jensen et al.).
  • Psychiatry and Behavioral Health: In sensitive domains like suicide risk assessment, Bhanu Pratap Singh Rawat et al. note that a model's inability to assert the correct temporal period prevents it from distinguishing a past suicide attempt from an immediate, active crisis.

Even state-of-the-art Large Language Models (LLMs) fall into this trap. Amey Hengle et al. evaluated LLMs on comorbid mental health diagnoses and found that they consistently over-diagnose patients. The models fail to recognize that a user describing a past traumatic experience is not necessarily exhibiting an active clinical symptom today.

Why Sentence-Level Context Fails

One might assume that modern transformer-based models, which utilize attention mechanisms to capture context, would naturally solve this problem. However, standard named entity recognition (NER) often operates at the sentence or short-token-window level.

As Aokun Chen et al. point out in their work on extracting delirium symptoms, sentence-level models lack the broader document context required to determine if a symptom is current or historical. A sentence like "The patient experienced severe agitation" might be nested under a section header titled "Past Medical History" or "Discharge Status from 2021." Without a mechanism to propagate that high-level temporal anchor down to the individual mention, the model treats the symptom as a present reality.

Furthermore, clinical text is filled with complex temporal dependencies and structural variations. Medications might be listed across different sections of a narrative note, making it difficult to determine if a drug is part of a patient's history or a newly intended action (Aman Sinha 0002 et al.). Similarly, documentation of a planned action (e.g., "will notify the attending physician") is frequently conflated by automated systems with a confirmed past action, violating critical safety auditing protocols (Paras Lakhani et al.).

Architectural Solutions: Decoupling and Memory

To resolve these ambiguities, researchers have shifted away from single-pass classification toward architectures that explicitly model time and history.

1. Decoupling Temporal Reasoning

Instead of asking a model to simultaneously identify a medical concept and determine its relevance, a more effective strategy is to separate the tasks. Zonghai Yao et al. demonstrate that decoupling temporal reasoning into a distinct classification stage significantly improves accuracy when extracting time-sensitive clinical conditions, such as social determinants of health like eviction risk. By isolating the temporal evaluation, the model avoids the common failure mode of misinterpreting historical mentions as immediate risks.

Naive Single-Pass Approach:
[Clinical Text] ──> [Single Model] ──> High False Positives (Confuses Past/Present)

Decoupled Approach:
[Clinical Text] ──> [Entity Extraction] ──> [Dedicated Temporal Classifier] ──> Accurate Active States

2. Infusing Models with "Memory" and Constraints

Another approach involves explicitly constraining models using a patient's historical baseline. Thomas Ferté et al. developed the PheVis algorithm, which incorporates a tunable "memory" of a patient's past medical history. This prevents the system from generating persistent positive predictions for acute conditions that have already been cured.

When generating radiology reports, integrating historical data helps models understand whether a condition is stable, worsening, or emergent, though care must be taken to prevent the model from hallucinating or failing to notice new clinical developments (HC-LLM).

3. Timeline Alignment and Coreference Resolution

To build a true longitudinal view of a patient, models must determine if two mentions across different notes refer to the same persistent condition or two distinct events. Eric Fosler-Lussier et al. utilize semantic and temporal features to resolve medical concept coreferences, helping systems differentiate between recurring acute events and persistent chronic conditions.

To stitch these events into a coherent timeline, the authors also leveraged Weighted Finite State Transducers (WFSTs) to align event sequences across multiple narrative notes, preventing the incorrect merging of separate medical incidents that happen to share similar relative time windows (Eric Fosler-Lussier et al.).

Beyond Diagnosis: Social and Administrative Context

The necessity of temporal state disambiguation extends far beyond tracking diseases; it is equally vital for understanding a patient's broader life circumstances.

Social Determinants of Health (SDOH)—such as housing instability, food insecurity, or employment status—are highly dynamic. Adam D. Lelkes et al. note that clinical notes frequently contain complex temporal information and contradictions regarding a patient's social history. If an NLP tool extracts a patient's occupation or housing status from a note, it must analyze temporal expressions to determine if that status represents past employment or current circumstances (Natasha Chilman et al.).

Ultimately, building clinical AI that truly understands a patient requires moving past simple keyword matching and entity extraction. By treating time as a foundational axis of interpretation, modern clinical NLP frameworks can accurately separate a patient's historical background from their immediate medical needs.

Key Foundational Work

Characterizing Non-Intermittent Faults (1991) — contributed 1 concept elements to this direction.
Automated Detection of Radiology Reports that Document Non-routine Communication of Critical or Significant Results (2010) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_temporal_state_disambiguation_9571a3fc,
  title = {temporal state disambiguation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/temporal-state-disambiguation/rdr_9571a3fc6840490f88a0c08ace394e39},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_temporal_state_disambiguation_9571a3fc_v1,
  title = {temporal state disambiguation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/temporal-state-disambiguation/rdr_9571a3fc6840490f88a0c08ace394e39/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake