A rising lactate or a falling blood pressure can mean very different things. The same number may reflect a drug the team just started, a disease getting worse, or a measurement nobody can yet explain. Clinicians usually record which of these they believe in their notes. They rarely spell it out, though. They write telegraphic runs like "Lactate 4.1. On epi. Will trend." The word "because" is missing, and the reader has to infer the cause. In linguistics, a link between two sentences that carries no connecting word is called an implicit discourse relation. This proposal treats those unstated links in serial progress notes as the missing explanation layer for longitudinal biomarker tracking. Longitudinal tracking means following lab values and vital signs over time instead of reading one snapshot. The idea is to detect each notable change in a patient's biomarker trajectory and find the note sentences that mention it. A model then infers the connective the clinician left out. That inference sorts each change into one of four groups: attributed to an intervention, attributed to the disease, acknowledged but discounted, or left unexplained.
The core bet is that a biomarker change and the clinician's stated explanation for it predict deterioration better together than either does alone. In particular, a change the clinician has not explained should carry more risk than the same-sized change pinned on a treatment. There is also a methodological twist. The structured timeline acts as a second "language" that partially makes the note's logic explicit. Medication timestamps, for example, show whether a drug started before or after a lab value moved. That lets us mine weak training labels for clinical discourse relations in the same spirit as mining connectives from human translations.
Core Claim
Existing systems that combine clinical text with structured data read notes as bags of words or flat token streams. Existing systems that track biomarkers read numbers without the clinician's interpretation. We claim three things that, to our knowledge, nobody has tested. First, implicit relations between note statements about a biomarker change can be recovered well enough to be useful, using connective generation, where a model predicts the missing linking word. Second, the resulting attribution of each change adds predictive value on top of both trajectory-only models and fused text-plus-structure models of equal capacity. Third, that added value comes specifically from the implicit cases, the ones a keyword rule for "due to" or "secondary to" would miss. If the third claim fails, the right conclusion is that explicit-marker rules suffice. The experiments are designed so that this outcome is visible.
Where This Sits
The biomarker literature motivates the focus on changes rather than levels. Serum sodium fluctuation predicts in-hospital mortality better than baseline sodium (Sodium fluctuation as a parameter in predicting mortality in general hospitalized patients). Albumin trajectories during hospitalization outperform admission albumin (Association between albumin changes and prognosis in older patients with acute myocardial infarction). Blood pressure variability predicts postoperative delirium even when the mean is normal (Perioperative blood pressure variability as a risk factor for postoperative delirium in the patients receiving cardiac surgery). On the modeling side, Optimal intensive care outcome prediction over time using machine learning shows that cumulative daily models beat a static admission severity score that loses predictive power after day one. Monitoring ICU Mortality Risk with a Long Short-Term Memory Recurrent Neural Network updates mortality risk from event sequences. All of these treat a change as a context-free signal. None asks whether the care team had already explained it.
A second group of work motivates bringing in text. In Creating an automated trigger for sepsis clinical decision support at emergency department triage using machine learning, adding bag-of-words nursing notes to vital signs lifts triage sepsis detection from AUC 0.67 to 0.86. Predicting ICU Length of Stay for Patients using Latent Categorization of Health Conditions reports that narrative text stabilizes length-of-stay forecasts early in the stay. Predicting Multimodal Modeling of Diagnoses and Treatments in EHR and M-BEHRT (Mbaye et al.) fuse notes with structured records as token sequences. These works show that text carries signal, but they never ask which relation links one note statement to another or to a lab change.
Horng et al. also constrain the design. Their linear model matched nonlinear ones because keyword presence drove performance. That is precisely the result a discourse model has to beat, not assume away. EHRSHOT is the strongest longitudinal foundation benchmark in the file, but it excludes notes entirely. It therefore cannot host this study, and it marks the text gap. Revisiting Clinical Outcome Prediction for MIMIC-IV benchmarks PubMedBERT and related encoders on single admission notes, with no serial structure.
The discourse literature supplies the method. Annotation-Inspired Implicit Discourse Relation Classification with Auxiliary Discourse Connective Generation inserts a mask between two sentences, predicts the missing connective, and classifies the relation from the arguments plus that connective. Training is end-to-end through a Gumbel-Softmax relaxation, a differentiable stand-in for picking one discrete word. On the Penn Discourse Treebank 2.0 top-level senses, this raises accuracy from 68.61% (plain RoBERTa) to 74.59%. We reuse this mechanism directly.
Using Explicit Discourse Connectives in Translation for Implicit Discourse Relation Classification mined silver training data from places where translators added connectives, raising 11-way accuracy from 34.32% to 45.50%. This is the template for our cross-modal mining. The parallel rests on evidence that translators make 25–30% of implicit relations explicit (Towards a discourse relation-aware approach for Chinese-English machine translation). We are betting that the structured timeline makes the temporal part of a note's logic explicit in a similar way.
DiscoGeM 2.0 supplies the annotation protocol. Annotators drag a connective into a slot, ten per item, which yields soft label distributions. It also gives a caution: agreement within a language is only moderate (Jensen-Shannon divergence 0.43), so clinical labels should be modeled as distributions, not gold classes. Two further papers constrain the design. Yangfeng Ji et al. show that text with its connectives stripped out differs from naturally implicit text, so we cannot simply delete "due to" from clinical notes and call the result implicit training data. I-Ta Lee et al. show that general language models lean on topical similarity rather than discourse role. That is the failure mode we expect from flat encoders.
We deliberately do not adopt full document discourse trees as in Disco2NMT. That approach depends on an upstream parser, the paper itself names parser noise as its bottleneck, and its gain in BLEU (a translation quality score) is small (24.19 to 24.60). Local relations anchored to biomarker changes are a narrower and more checkable target.
Research Program
The study uses MIMIC-IV, which pairs timestamped labs, vitals, and medication administrations with progress and nursing notes. Preprocessing follows An Extensive Data Processing Pipeline for MIMIC-IV where possible.
The pipeline has four stages.
Stage one: detect changes. We detect excursions in roughly ten high-volume biomarkers: lactate, creatinine, sodium, potassium, hemoglobin, white count, mean arterial pressure, heart rate, oxygen saturation, and temperature. An excursion is a departure from the patient's own recent trend, not from a population reference range. This follows the trajectory-over-level lesson above.
Stage two: link changes to note text. Each excursion is linked to note sentences written within a window after it. The link uses mention matching on the biomarker name, its abbreviations, and its values. The linked sentence becomes one argument. Its neighboring sentences, plus a templated sentence describing nearby structured events such as "norepinephrine started 02:10," become the candidate second arguments.
Stage three: infer the missing connective. The connective-generation classifier of Liu and Strube is initialized from a biomedical encoder and trained in three phases. It first learns from the Penn Discourse Treebank. It then trains on cross-modal silver data. Finally it trains on a small clinician-annotated set.
The silver data comes from note sentence pairs where the clinician did write a connective, and where the structured record independently agrees with that connective's temporal direction. A case like "MAP improved after starting levophed" qualifies when the timestamps show the drug came first. These pairs serve as explicit-anchored examples. We then add naturally implicit pairs whose structured timeline makes one reading strongly preferred. Following Ji et al., we treat explicit-derived and naturally implicit data as separate domains rather than pooling them.
The annotated set covers about 2,000 implicit sentence pairs about excursions. Each is labeled by several clinicians using DiscoGeM's drag-a-connective interface with a small clinical connective inventory: because of treatment, because of disease, despite, then, and no relation. The output is a soft distribution per pair.
Stage four: predict outcomes. Every excursion receives an attribution distribution over the four categories: intervention-explained, disease-explained, discounted, and unexplained. A time-varying risk model uses these attributions alongside the raw trajectory. The model is the same sequence architecture as the strongest trajectory baseline, plus attribution channels. This setup keeps the comparison about the attribution information, not about model size. Outcomes are in-hospital mortality, deterioration within 48 hours (ICU transfer or new vasopressor), and acute kidney injury onset. Predictions are updated daily, following the cumulative design of Meiring et al.
Baselines
The naive reference point is a static admission model. It is a logistic regression on first-24-hour vitals and labs, which approximates the admission severity scoring that Meiring et al. show decays after day one. Any longitudinal model should beat it easily. Its role is to calibrate how much of the total gain comes from tracking at all.
The strongest structured-only competitor is a daily cumulative trajectory model in the style of Optimal intensive care outcome prediction over time using machine learning and Monitoring ICU Mortality Risk with a Long Short-Term Memory Recurrent Neural Network. We will also run a gradient-boosted tree model on trend and fluctuation features, because EHRSHOT found count-based boosted trees competitive with a 141M-parameter foundation model at larger sample sizes. On a different ICU database (CCHIC), Meiring et al. report day-1 mortality AUC 0.883 for their deep network versus 0.836 for the admission score, rising to 0.895 on day 2. These numbers are context only, not a target on our cohort. Both models are retrained on our exact MIMIC-IV cohort, splits, and daily landmarks.
The strongest text-aware competitor is a fused model that concatenates serial notes into a PubMedBERT encoder and joins that encoding with the trajectory model. PubMedBERT was the best encoder in Revisiting Clinical Outcome Prediction for MIMIC-IV, which reports ED diagnosis prediction AUROC 86.86% and PR-AUC 17.24%. That is a different task, and we found no reported result on Lacuna for note-plus-trajectory deterioration or mortality prediction on MIMIC-IV.
We add a bag-of-words linear model over the same notes, because Horng et al. found that this simple design matched nonlinear models. On their triage sepsis task it reached AUC 0.86 with vitals plus notes. All text-aware baselines see exactly the same notes within the same time windows, use the same encoder size where applicable, and get the same hyperparameter search budget, measured in trials, as our model.
Two ablations isolate the key ingredient. The first, an explicit-only attribution model, assigns attributions with rules over written connectives and phrases such as "due to," "secondary to," and "despite," and leaves implicit cases unexplained. The second keeps the full pipeline but removes the connective-generation step, so attribution comes from a plain classifier over the argument pair. For the intrinsic discourse task, the reference point is Liu and Strube's reported 74.59% top-level accuracy on the Penn Discourse Treebank 2.0 versus 68.61% for plain RoBERTa. For clinical notes, we found no reported result on Lacuna for implicit discourse relation classification.
Beating the baselines means an improvement in AUROC and in PR-AUC over the fused PubMedBERT baseline at matched capacity, with patient-level bootstrap intervals that exclude zero, on at least two of the three outcomes. It also means a gain over the explicit-only ablation, concentrated in excursions whose linking sentences contain no explicit connective.
Why not just…?
Why not just feed all serial notes and labs into a clinical transformer and let attention find the relations? This is the main threat, and we take it seriously. Lee et al. found that general language models track topical similarity rather than discourse role. Liu and Strube found that an explicit connective step adds about six points over plain RoBERTa even on clean newswire. Neither finding was made on clinical notes, however. The fused PubMedBERT baseline, run at matched capacity, settles the question directly. If it matches our model, the attribution layer adds nothing beyond interpretability, and we will report that. A probing check adds diagnostic detail. We freeze the fused model's note representations and test whether a linear probe can recover the clinician-annotated attribution labels as well as our classifier does. If it can, the information was already there and our pipeline is redundant.
Why not just use keyword rules for explicit causal phrases? If clinicians usually write "due to," rules win on cost. The explicit-only ablation answers this. We will also report what fraction of excursion-linked sentence pairs carry no explicit marker. If that fraction is small, or if the explicit-only ablation matches the full model, the implicit-relation machinery is not justified for this problem. We would then reframe the contribution as a rule-based attribution feature.
Why not just model the trajectory more carefully, using fluctuation, slope, and interaction features, and skip text? The biomarker literature itself suggests that well-engineered trend features are strong. Text added large gains in Horng et al. and earlier stabilization in Dasgupta et al. Neither study tested whether text helps once trajectory features are rich, so this question is open. The gradient-boosted tree and sequence baselines carry rich fluctuation features. The stratified analysis below tests whether attribution helps exactly where trajectories are ambiguous: same-sized excursions with different clinical explanations.
Why not just ask a large language model to label each excursion's cause zero-shot? This is a reasonable labeling alternative, and we will include it as a fourth attribution source scored against the clinician-annotated set. It does not replace the cross-modal silver data, which grounds labels in timestamps rather than plausible-sounding text. It also does not replace a small classifier that can run over millions of note pairs. If the language model's attributions match ours in both intrinsic agreement and outcome gain, we will adopt it and say so.
Decisive Checks
The first check is intrinsic. We compare the connective-generation classifier with the no-connective variant and the zero-shot language model on the clinician-annotated set. The metric is Jensen-Shannon divergence to the soft label distributions, plus top-label accuracy. Interpretation is calibrated against the inter-annotator divergence, because DiscoGeM 2.0 shows that even human agreement is limited.
The second check validates the silver data. We hold out timestamp-verified pairs and measure whether training on cross-modal silver data improves accuracy on human-labeled naturally implicit pairs. If silver data helps only on explicit-derived pairs, the Ji et al. domain gap has bitten, and the mining idea fails.
The third check is the main extrinsic comparison of daily AUROC and PR-AUC against all baselines above. It includes a lead-time analysis of how many hours earlier each model crosses a fixed alert threshold before deterioration, echoing the text lead-time argument of Dasgupta et al.
The fourth check is the mechanism test. We match excursions on biomarker type and magnitude, then compare outcome rates across attribution categories. The prediction is that unexplained and disease-attributed excursions carry higher subsequent risk than intervention-attributed ones of the same size. If the categories do not separate risk, the core hypothesis is wrong regardless of aggregate AUROC.
The fifth check is a placebo test: shuffling attributions across excursions within a patient should erase the gain. If the gain survives shuffling, the model is exploiting note volume or timing rather than discourse content.
References
- Sodium fluctuation as a parameter in predicting mortality in general hospitalized patients
- Association between albumin changes and prognosis in older patients with acute myocardial infarction
- Perioperative blood pressure variability as a risk factor for postoperative delirium in the patients receiving cardiac surgery
- Optimal intensive care outcome prediction over time using machine learning
- Monitoring ICU Mortality Risk with a Long Short-Term Memory Recurrent Neural Network
- Creating an automated trigger for sepsis clinical decision support at emergency department triage using machine learning
- Predicting ICU Length of Stay for Patients using Latent Categorization of Health Conditions
- Predictive Multimodal Modeling of Diagnoses and Treatments in EHR
- EHRSHOT: An EHR Benchmark for Few-Shot Evaluation of Foundation Models
- Revisiting Clinical Outcome Prediction for MIMIC-IV
- Annotation-Inspired Implicit Discourse Relation Classification with Auxiliary Discourse Connective Generation
- Using Explicit Discourse Connectives in Translation for Implicit Discourse Relation Classification
- Towards a discourse relation-aware approach for Chinese-English machine translation
- DiscoGeM 2.0: A Parallel Corpus of English, German, French and Czech Implicit Discourse Relations
- Closing the Gap: Domain Adaptation from Explicit to Implicit Discourse Relations
- Multi-Relational Script Learning for Discourse Relations
- Towards Discourse-Aware Document-Level Neural Machine Translation
- An Extensive Data Processing Pipeline for MIMIC-IV