Many of the events that matter most in an intensive care stay are written down in prose before, or instead of, appearing in structured tables. Delirium, a new infection and the first signs of sepsis are typical examples. A trauma-ICU forecasting model built as a dynamic Bayesian network had to leave these complications out for exactly this reason. A dynamic Bayesian network is a probabilistic model that tracks hidden patient states from one time step to the next and updates its beliefs as evidence arrives. The authors say the onset of these complications is recorded mainly in nursing narratives (Dynamic Bayesian network for predicting physiological changes, organ dysfunctions and mortality risk in critical trauma patients). The obvious fix is to run a text classifier over each note and feed its outputs into the network as evidence. This proposal argues that the obvious fix is miscalibrated in a predictable way. Clinical notes are not independent sensor readings. Clinicians copy text forward from one note to the next, so a finding written once can reappear across ten notes and look like ten confirmations. Notes also describe the past as often as the present ("rash resolved last week", "s/p biopsy"). A model that treats every mention as fresh, current-time evidence becomes overconfident and puts onsets at the wrong time.
We propose to model the documentation process itself. Each note-level extraction becomes an observation of the hidden state at some earlier, uncertain time. It passes through a learned error model that depends on assertion type (present, negated, hedged, historical). It also includes an explicit "copied forward" path, so that repeated text is explained by the previous note rather than counted again. The central claim is this: on real MIMIC-III notes, a network with this documentation layer tracks the onset of text-documented complications with better-calibrated beliefs and smaller timing error than two alternatives. The first is the same network with naive per-note evidence injection. The second is an end-to-end neural model reading the same notes. The documentation layer also keeps the auditability that makes graphical models attractive to clinicians.
Core Claim and What Is New
The closest existing work is Clinical Reasoning over Tabular Data and Text with Bayesian Networks. It trains separate text classifiers on clinical-language embeddings, injects their predicted symptom probabilities into a static Bayesian network, and reaches an average precision of about 0.75, against about 0.66 for a neural network that concatenates all features. That paper motivates this proposal and supplies its basic mechanism. It is also static, runs on 5,000 synthetic GPT-3.5 notes, and has exactly one note per patient, so copy-forward, historical mentions and evidence accumulation over time never arise.
The same paper also constrains our design. Its generative variant modeled text embeddings as Gaussians conditioned on disease state, and it failed completely. We keep a generative story but apply it one level up. The network does not generate embeddings. It generates the extractor's discrete outputs, through a small error model with a time offset. That model has few parameters and does not need to fit a high-dimensional embedding space.
The dynamic side supplies the time structure and the evaluation cohort. The trauma network is learned with the PC algorithm (a method that infers edges from conditional independence tests) under expert constraints, and it reports a 24-hour mortality AUC of 0.977 on MIMIC-III (Dynamic Bayesian network for predicting physiological changes, organ dysfunctions and mortality risk in critical trauma patients). We reuse its template of two time slices and extend it with latent complication nodes observed only through text.
The Documentation Observation Model
Let $Z_t$ be the hidden complication state in time slice $t$, for example whether delirium is active. Let note $n$ be written in slice $\tau_n$, and let $m_n$ be the extractor's label for that complication in the note: present, absent, hedged, historical, or not mentioned. We model
$$P(m_n \mid Z_{1:\tau_n}, m_{n-1}) = \rho_n\, \delta(m_n = m_{n-1}) + (1-\rho_n) \sum_{k \ge 0} \pi(k \mid a_n)\, C(m_n \mid Z_{\tau_n - k}).$$
In this equation, $\rho_n$ is the probability that the mention was carried forward from the previous note. It is predicted from lexical overlap between consecutive notes of the same type, so verbatim copies push $\rho_n$ toward one. The term $\pi(k \mid a_n)$ is a distribution over how many slices back the mention refers. It depends on a coarse temporal cue $a_n$ extracted from the sentence, such as tense, "resolved" or "s/p". Finally, $C$ is a confusion matrix linking true state to extracted label.
"Not mentioned" is a real observation with its own row in $C$. In routine charting, silence about delirium is weak evidence against it, and the model should learn how weak. Because the offset $\pi$ is inferred jointly with the state sequence during smoothing, the extractor does not have to fix exact event times. This matters because Investigating the Challenges of Temporal Relation Extraction from Clinical Text finds language models struggle with exactly that for nominal events such as "biopsy". Physiology evidence in neighbouring slices can do some of that work instead.
Research Program
We build on the MIMIC-III critical trauma cohort defined as in the trauma-network paper and keep its structured variables and slice width. We add latent nodes for three complications: hospital-acquired infection, sepsis and delirium. They connect to the existing physiology and organ-dysfunction nodes by edges allowed under expert constraints. Transitions within a slice and across slices are learned as in the original model, which keeps structure learning familiar and puts the novelty in the observation layer.
The text extractor is a fine-tuned clinical transformer that labels each note with complication mention and assertion status. It uses entity-span context pooling, which MIMICause showed raised causal-relation macro F1 from 0.23 to 0.54 over pooled-sentence baselines. We borrow the pooling technique, not the task.
Training labels come from the hybrid recipe of CheXbert: rule-based labels at scale, corrected by a few hundred expert-annotated notes. The parameters $C$, $\pi$ and the copy-forward model are fitted by expectation–maximization over the network, holding the extractor fixed. Exact inference stays tractable because each complication node has a small number of states, and the observation layer adds only local factors per note.
The ground-truth design reuses the masking idea of the static hybrid paper on real notes. Onset times come from structured sources that are hidden from every model at test time. Infection onset uses microbiology culture positivity, the same source that underlies the FTL-Trans infection tasks (Time-Aware Transformer-based Network for Clinical Notes Series Prediction). Sepsis onset uses a suspected-infection criterion (culture draw plus antibiotics) combined with organ dysfunction. Delirium uses charted delirium screening assessments where they exist. The models therefore have to recover onset from text plus the remaining physiology. The comparison is fair because every system sees the same masked inputs.
Baselines
The naive reference point is the published trauma network run without text. It cannot represent the masked complications, so its complication beliefs are simply the population prior carried forward, and it serves as the floor. On mortality it anchors the upper end: its reported 24-hour mortality AUC is 0.977 on MIMIC-III. We do not expect text to improve that figure much, and we would treat any large drop as a sign that we broke the structured model.
The key ablation is the same dynamic network with naive injection. Each note's calibrated classifier probability enters as independent current-slice evidence, which extends BN-discr-text directly into time. It shares the extractor, structure, data split and inference code with our model and differs only in the observation layer. It is the most honest test of whether the documentation model matters. Two narrower ablations remove only copy-forward (setting $\rho_n = 0$) and only the offset (placing all mass on $k=0$), so we can attribute any gain to each part. Beating this ablation means lower median onset-timing error, lower Brier score for per-slice complication beliefs, and passing the duplication test described below, each with bootstrap confidence intervals that exclude zero.
The strongest neural competitor is a time-aware sequence model over notes and vitals. We use FTL-Trans (Time-Aware Transformer-based Network for Clinical Notes Series Prediction), which handles irregular gaps between notes with a learned time decay and reports 95.00% AUROC for in-hospital mortality. We give it a structured-vitals stream in the style of Predicting in-hospital mortality by combining clinical notes with time-series data, whose fused model reports an AUC of 0.899. We train it to emit a per-slice onset hazard for each complication under the same masking, with the same tuning budget in GPU-hours and hyperparameter trials. On onset timing and slice-level calibration, no reported result was found on Lacuna for any of these models, so we establish those numbers ourselves. Beating this competitor means matching its onset discrimination within confidence intervals while showing clearly better calibration and duplication robustness. Winning on discrimination is not required for the claim.
Why not just…?
Why not just use a strong multimodal neural model and skip the graphical model? Neural fusion already predicts mortality well: 0.899 AUC in the fused note-plus-vitals model and 0.8787 AUROC for MM-HCR at 24 hours (Towards unstructured mortality prediction with free-text clinical notes). The authors of both that work and the M3 benchmark (A Multi-Modal and Multitask Benchmark in the Clinical Domain) name opacity and the lack of latent causal factors as the open problem. Our question is different: a calibrated, inspectable belief about when a specific complication began. The neural baseline above is trained directly for that target. If it matches us on calibration and duplication robustness as well as discrimination, the case for the graphical model reduces to interpretability, and we will report that plainly.
Why not just feed per-note classifier outputs into the existing network? This is the naive-injection ablation, and it is the real threat to the proposal. If copy-forward and historical mentions turn out to be rare in trauma-ICU nursing notes, the documentation layer will add parameters without adding accuracy. The duplication test and the ablation on the real cohort settle this directly. We will also report how often consecutive-note copy overlap occurs, so readers can see whether the phenomenon is large enough to matter.
Why not just have a long-context language model read the whole stay and state the onset time? Models such as ClinicalMamba handle 16,000-token records, and Dementia-R1 shows that reinforcement learning on intermediate verifiable quantities improves longitudinal reasoning. Such a model gives one answer, though, not a posterior that updates slice by slice, and its stated confidence is not known to be calibrated. We include a prompted long-context model that outputs onset slice and confidence as a secondary baseline. We measure its timing error and calibration under the same masking, and we note that it may well beat us on timing alone.
Proposed Experiments
The primary experiment runs on the MIMIC-III trauma cohort with patient-level train, validation and test splits, and with structured onset signals hidden at test time. For each complication we report four things: median absolute onset-timing error in hours; Brier score and expected calibration error of per-slice beliefs; sensitivity at a fixed alarm rate per patient-day; and, as a secondary check, 24-hour mortality AUC against the published 0.977.
The decisive check is a duplication intervention. In test stays we insert verbatim copies of existing notes, and separately rewrite some present-tense mentions as historical ones. A correctly specified model should barely change its posterior after a pure duplicate and should move onset estimates earlier after a historical rewrite. The naive-injection model should become overconfident under duplication. If our model's belief shift under duplication is not substantially smaller than naive injection's, the central claim fails.
A second check asks whether the learned offset distribution $\pi$ tracks linguistic cues sensibly, for example whether "resolved" mentions point further back than hedged ones. We audit this on a few hundred annotated sentences. A third check applies the fitted observation layer to psychiatric symptom tracking, in the spirit of the survey-based network of Links between psychotic and neurotic symptoms in the general population, only if suitable note data can be obtained. We treat it as optional and not required for the main result.
References
- Dynamic Bayesian network for predicting physiological changes, organ dysfunctions and mortality risk in critical trauma patients
- Clinical Reasoning over Tabular Data and Text with Bayesian Networks
- Investigating the Challenges of Temporal Relation Extraction from Clinical Text
- MIMICause: Representation and automatic extraction of causal relation types from clinical notes
- Combining Automatic Labelers and Expert Annotations for Accurate Radiology Report Labeling Using BERT
- Time-Aware Transformer-based Network for Clinical Notes Series Prediction
- Predicting in-hospital mortality by combining clinical notes with time-series data
- Towards unstructured mortality prediction with free-text clinical notes
- A Multi-Modal and Multitask Benchmark in the Clinical Domain
- ClinicalMamba: A Generative Clinical Language Model on Longitudinal Clinical Notes
- Dementia-R1: Reinforced Pretraining and Reasoning from Unstructured Clinical Notes for Real-World Dementia Prognosis
- Links between psychotic and neurotic symptoms in the general population: an analysis of longitudinal British National Survey data using Directed Acyclic Graphs