Denied, Not Missing: Negated Findings as Time-Aligned Signals for Biomarker Trajectories

Clinical notes spend much of their length saying what a patient does not have: "no signs of fluid overload," "denies chest pain," "infection ruled out." Text-mining pipelines treat these statements in one of two ways. Some throw them away, the classic job of negation filtering. Others feed them to a neural encoder and hope it reads the "no" correctly. Neither approach notices that a negated finding is itself a dated clinical event. A clinician wrote it because the condition was being considered, and it was written at a specific moment while the patient's labs and vital signs were moving. We propose to turn each patient's notes into a timeline of concept-level assertion states: affirmed, denied, uncertain, historical, or simply not mentioned. These timelines sit on the same hourly grid as serial lab and vital-sign measurements. The test is whether keeping "denied" as its own state, aligned in time with the direction and rate of biomarker change, predicts ICU deterioration better than deleting negations or encoding raw text.

The opportunity matters because the two literatures have stayed separate. Longitudinal biomarker models track how fast a measurement changes and in which direction, but they ignore the notes. Negation-aware text systems are almost all single-snapshot models. The few systems that fuse notes with ICU time series either average all notes into one vector or down-weight older notes by recency, and none model negation. Our claim is narrow and testable. When a clinician denies a finding while a related biomarker is getting worse, that disagreement is a warning sign. A model can only see it if negation is represented explicitly and placed on the timeline.

Core Claim

Negation filtering in its standard form, as in NegEx-based sepsis triage or the cTAKES COPD pipeline, either deletes negated mentions or tags them once per document. We argue that in a longitudinal setting, three states differ and should stay separate: "denied at hour 20," "affirmed at hour 20," and "never mentioned." Changes between these states, together with the velocity of related biomarkers, carry information that neither stream carries alone.

The ingredient that separates this from prior work is a five-state assertion trajectory per concept, aligned to the biomarker time grid. The proposal stands or falls on one ablation. Collapsing "denied" into "unmentioned" reproduces classic filtering. Collapsing "denied" into "affirmed" reproduces no filtering. If neither collapse costs performance, the claim is wrong.

Research Program

The work builds directly on the two fusion systems closest to this question.

Deznabi et al. encode 48 hours of vitals with an LSTM and fine-tuned Bio+Clinical BERT notes, then mean-pool the notes into a single vector. That pooling erases when each statement was made.

Khadanga et al. use Word2Vec and a convolutional network over notes, weighting each note by an exponential decay in its age. Their encoder represents "no dyspnea" and "dyspnea" with the same symptom tokens.

Both systems motivate the gap. Neither represents polarity explicitly, and neither aligns individual statements with biomarker inflections.

Assertion timeline construction. From MIMIC-III nursing and physician notes timestamped within the first 48 ICU hours, we extract UMLS concepts and assign each mention one of the five states. The assertion tagger is pluggable, and we compare three versions:

  • an extended NegEx rule set with section-header attribution, so that mentions under "Family History" or "Past Medical History" are labeled historical or familial;
  • a RoBERTa scope model with negation-focused pre-training, following CueNB, which reported cross-dataset gains of +3.31 F1 for cue detection and +1.72 F1 for scope resolution;
  • a prompted large language model extractor.

Because CueNB documents inconsistent annotation conventions across negation corpora, we do not trust published tagger scores to transfer. We will hand-annotate a few hundred MIMIC sentences to measure tagger accuracy directly. This is especially important for enumerated lists and constructions like "no X other than mild Y," which that paper flags as failure cases.

Shared time grid. We restrict to the K most frequent concepts that clinicians link to deterioration, roughly 100 to 200, covering infection, respiratory, cardiac, and fluid status. Each concept becomes a channel on the MIMIC-Extract hourly grid. A channel holds the most recent assertion state plus the time since that state was set. This mirrors how irregular-sampling models already treat missing labs: the absence of a measurement is itself informative.

Prediction model. A single recurrent model with missingness-aware inputs consumes the lab and vital channels together with the assertion channels. We add one small, interpretable component: learned interaction terms between each concept's state and the local slope of a short list of related biomarkers. Examples are infection terms paired with neutrophil-to-lymphocyte ratio (NLR) and lactate, and fluid-overload terms paired with albumin and creatinine. The model is deliberately simple. If gains appear, they should be attributable to the representation and not to model capacity.

Analysis target. Salciccioli et al. found that admission NLR predicts 28-day mortality across ICU patients (adjusted OR 1.71 for the top quartile) but loses significance in sepsis, where the ratio saturates. We will test whether assertion trajectories restore discrimination inside that sepsis subgroup. This is where biomarkers alone are known to fail, and it is the most demanding test of whether text contributes something the numbers cannot.

Baselines

Naive reference. The do-nothing baseline is a structured-only model: the same recurrent architecture on MIMIC-Extract labs and vitals, with no text. Deznabi et al. report AUC 0.835 for their time-series-only LSTM on 48-hour in-hospital mortality. MIMIC-Extract reports mortality AUROC between 0.85 and 0.90 depending on model. We retrain this baseline on our split instead of copying those figures, because cohort definitions differ.

Strongest existing method. The strongest competitor is the Deznabi multimodal model: fine-tuned Bio+Clinical BERT on each note, mean-pooled, concatenated with an LSTM over vitals. It reports AUC 0.899 for in-hospital mortality versus 0.875 for text alone. It is the strongest because a fine-tuned transformer may already resolve most negation implicitly, which is exactly the hypothesis we must defeat.

We also run the Khadanga decay-weighted fusion. It reports mortality AUPRC 0.525 (versus 0.487 for time series alone) and decompensation AUPRC 0.345 (versus 0.325).

All baselines are reimplemented on one shared patient-level split. They use the same note-time cutoff, which excludes discharge summaries and any note charted after the prediction time. They use the same lab channels and receive equal hyperparameter-search budgets. We also give the BERT baseline a stronger time-aware variant: per-window note embeddings fed into the recurrent model in place of a single mean. This way any gain we show cannot be attributed simply to undoing mean-pooling.

Ablations. The ablations isolate the key ingredient. "Denied→unmentioned" corresponds to classic filtering. "Denied→affirmed" corresponds to no filtering. A third variant keeps the states but removes their timing, using one per-stay bag of states.

What counts as beating the baselines. We define success before running anything: an AUPRC improvement over the time-aware BERT fusion whose patient-level bootstrap 95% interval excludes zero, together with a significant loss when "denied" is collapsed. An AUC gain alone will not count, since AUC is already near ceiling at roughly 0.9.

Why not just…?

Why not just fine-tune ClinicalBERT on the notes and let attention handle negation? This is the real threat. Deznabi's text-only model already reaches 0.875 AUC. CueNB shows standard encoders lean on shallow cues and degrade across domains, and the FORTE radiology evaluation had to strip negations by hand because general representations confused present and absent findings. Neither result shows that an ICU outcome model is actually hurt. We will measure this directly. A polarity-flip test edits negation cues in held-out notes and checks whether the dense model's risk moves in the clinically correct direction. If the dense model already responds to polarity and matches our AUPRC, we will report that explicit assertion trajectories are unnecessary for this task.

Why not just apply standard NegEx and drop negated mentions? That is the "denied→unmentioned" ablation, and it is the established practice behind results such as Sontag et al., where triage AUC rose from 0.67 to 0.86, and the incontinence pipeline, with PPV 0.73 and sensitivity 0.84. If dropping negations performs as well as keeping them as a state, the core claim fails.

Why not just use labs and vitals, which already carry the trajectory? Khadanga et al. and Deznabi et al. both found that notes add to time series. The NLR collapse in sepsis reported by Salciccioli et al. is a concrete case where numbers saturate. The sepsis subgroup experiment tests whether text helps exactly there.

Why not just read the whole record with a long-context text model? ClinicalMamba handles up to 16,000 tokens of notes. However, it uses no lab series, and the file reports no ICU mortality result for it on Lacuna. Khadanga et al. report that text-only models underperform time series on ICU management tasks. We include a long-context text-only arm as a reference, not as a competitor we expect to lose to.

Proposed Experiments

The primary tasks are MIMIC-III 48-hour in-hospital mortality and hourly decompensation, scored by AUPRC and AUROC on the shared split. A secondary task forecasts whether lactate, creatinine, or NLR will rise or fall over the next 24 hours. No reported result was found on Lacuna for this forecasting task, so we report our own baselines only.

The decisive checks are four:

  1. Collapse ablations. Collapsing "denied" in either direction must cost AUPRC beyond bootstrap noise.
  2. Polarity-flip sensitivity. For each model, we measure how much predicted risk changes when negation cues in held-out notes are flipped. This shows whether the dense baselines ignore polarity.
  3. Sepsis subgroup. Assertion-plus-trajectory features should restore mortality discrimination where NLR alone fails.
  4. Discordance analysis. Stays in which a concept is denied while its paired biomarker worsens should show elevated observed deterioration, and the learned interaction terms should be largest for those pairings.

Tagger quality enters every result. We therefore repeat all four checks with the rule-based, CueNB-style, and language-model taggers, to confirm that conclusions do not depend on one extractor. A negative result from checks 1 and 2 together would be informative on its own: it would mean modern fused encoders have already absorbed what negation filtering was built to supply.