After prostate cancer surgery, patients recover urinary, sexual and bowel function slowly and unevenly, over many months. Almost none of this is recorded in a form a model can use. Billing codes miss it almost entirely. In one prostatectomy cohort, fewer than 0.1% of patients carried an incontinence code, yet 30.3% had incontinence written in their clinicians' notes (Hernandez-Boussard et al.). Validated patient questionnaires such as the Expanded Prostate Cancer Index Composite (EPIC, a multi-domain survey of urinary, bowel, sexual and hormonal function) measure recovery directly, but they are given only every six months or so, and patients stop answering them. In a large prospective prostate cancer study, only 54.8% completed the 12-month survey (READT). So the data come as two streams that are each broken in a different way. Surveys are accurate but sparse and prone to dropout. Notes are frequent but reflect what the clinician chose to write down.
We propose one model that treats both streams as noisy readings of the same hidden, multi-domain functional state. The state evolves in continuous time, and its recovery rate is conditioned on the tissue pathology found at surgery. Surveys enter as calibrated measurements. Notes enter as biased measurements whose silences the model learns to interpret: a note that says nothing about incontinence is not evidence that the patient is continent. The payoff is a per-patient forecast of recovery in each domain, including for the many patients who drop out of survey follow-up. That is the group current analyses discard or impute naively.
Core Claim
The claim is narrow and testable. When clinician notes are modeled as a domain-specific, bias-aware measurement channel on a latent functional state, rather than as labels or as extra covariates, they improve prediction of held-out patient-reported scores in each EPIC domain. The gain should be concentrated in patients whose surveys are missing.
Conditioning the recovery dynamics on surgical pathology is the second ingredient. The pathology features are margin status, pathological stage, extraprostatic extension and grade. We test this ingredient separately and claim it only if the ablation supports it.
Histopathology plays a particular role here. The field treats it as ground truth for diagnosis (Da Silva et al.; CMMD). We use it as the fixed baseline that sets each patient's recovery dynamics. This addresses the open question in the file of how a single cross-sectional tissue sample should condition a trajectory.
Relation to Prior Work
Three lines of work motivate the proposal.
- The documentation gap. Hernandez-Boussard et al. extracted incontinence mentions from notes with a dictionary tagger plus negation handling (PPV 0.73, sensitivity 0.84). They produced one binary label per patient within 90 days. We keep their extraction idea but turn mentions into timed, noisy observations.
- Multi-domain recovery is real and diverges across treatments. Ng et al. found 70% incontinence after robotic surgery versus 3.2% after radiotherapy, measured cross-sectionally at 5–7 years. Lourenço et al. validated a five-domain EPIC-CP structure at three time points but analyzed it only with paired tests. READT fit quantile mixed models to a single utility score.
- Single scores hide domain effects. Cong et al. showed domain subscores predict utility better than totals. This is the reason our latent state is a vector, not a scalar.
Three other works supply the method, and all three constrain the design.
- The deep state-space framework learns latent transitions from lab and vital-sign time series in chemotherapy patients.
- DySurv produces dynamic survival curves with a conditional variational autoencoder.
- Beyond Thresholds shows that trajectory features raise prolonged-ICU-stay AUPRC from 0.244 to 0.339, with the largest gains among patients labeled low risk.
These models ingest only objective physiology, target mortality or ICU outcomes, and assume their input channels are unbiased. Our model differs on all three counts: its observations are subjective, its targets are functional, and one of its channels is known to be biased.
One cited result bounds how much we can rely on text. The LoRA-tuned Llama PRO study reached only AUC 0.750 for depression from questionnaire text. That tells us text-derived signals are weak on their own and should be modeled as noisy measurements, not treated as truth.
Research Program
Each patient $i$ has a latent state $z_i(t) \in \mathbb{R}^D$, one coordinate per EPIC domain ($D=5$), in continuous time from surgery. The state drifts toward a patient-specific plateau:
$$dz_i = \Lambda(h_i)\,(\mu(h_i, x_i) - z_i)\,dt + \sigma\, dW_t,$$
The symbols are defined as follows:
- $h_i$ is the surgical pathology vector.
- $x_i$ holds baseline covariates: age, baseline EPIC scores, nerve-sparing status.
- $\Lambda$ is a diagonal matrix of recovery rates, one per domain.
- $\mu$ is the plateau, the level each domain recovers toward.
- $\Lambda$ and $\mu$ are small neural networks.
- $W_t$ is Brownian noise, so trajectories vary randomly around the drift.
This mean-reverting form is chosen because recovery curves after prostatectomy rise and then flatten. It also keeps the model interpretable: pathology changes either how fast a patient recovers or how far.
Survey channel. A survey at time $t$ is modeled as a Gaussian reading of $z_i(t)$ through the validated factor loadings.
Note channel. For domain $d$, at each clinical encounter a note either affirms a problem, denies it, or is silent. We model this as a three-way categorical outcome:
$$P(m_{i,d,t} = k) = \mathrm{softmax}_k\big(\alpha_{d,k} + \beta_{d,k}\, z_{i,d}(t) + \gamma_{d,k}^\top c_{i,t}\big),$$
The symbols are defined as follows:
- $m_{i,d,t}$ is the note label for patient $i$, domain $d$, time $t$.
- $k$ indexes the three outcomes: affirmed, denied, silent.
- $\alpha$, $\beta$, $\gamma$ are learned parameters.
- $c_{i,t}$ describes the encounter: urology versus other specialty, and visit type.
The model can therefore learn that sexual function is rarely mentioned outside urology visits. Silence then carries little information in those contexts.
Dropout. The probability that the next survey is missing depends on the current latent state. This lets us test whether dropout is informative, for example whether patients who recover stop responding.
Estimation and extraction. The model is fit by variational inference with an irregular-time encoder over both streams. Note labels come from an extractor with two options: the dictionary-plus-negation pipeline of Hernandez-Boussard et al., or a LoRA-tuned small language model in the style of the depression study. The extractor is validated on a manually annotated sample of note snippets per domain before any trajectory modeling.
Data. No resource in the file contains all three ingredients: repeated EPIC surveys, full clinical notes, and surgical pathology. The study therefore needs an institutional prostatectomy cohort of the kind described in Lourenço et al., with 402 patients surveyed pre-surgery and at 6 and 12 months. That cohort must be linked to the notes and pathology reports in the EHR, as in the Stanford cohort. We will say plainly that the public datasets in the file cannot host the main experiment. Public data are used only for component checks.
Baselines
Naive reference. Two naive predictors are run. The first carries forward each patient's last observed EPIC domain score. The second uses the domain-wise population mean at that time point. Every method must beat these on held-out survey scores.
Strongest existing method. This is a per-domain linear quantile mixed model with random patient intercepts and slopes and pathology as fixed effects. It is the analysis READT applied to repeated quality-of-life measures, extended to five domains. It is the real competitor, because recovery curves are smooth and mixed models handle sparse repeated measures well.
We found no published predictive accuracy on Lacuna for EPIC domain trajectory forecasting by any method. The closest reported numbers are for other tasks:
- the domain-level mapping error of Cong et al., MAE 0.1206 on a utility scale;
- the static-versus-trajectory AUPRC of 0.244 versus 0.339 in Beyond Thresholds.
The mixed model therefore gets its baseline number from our own run.
Note-covariate mixed model. This baseline gives the mixed model the same note information, as time-varying counts of affirmed and denied mentions per domain.
Deep state-space without notes. This is our architecture with the note channel removed. It follows the spirit of the deep state-space framework.
Fairness. All methods use the same patients, folds and pathology and baseline features. Each gets the same hyperparameter search budget, and each is tuned on a validation split that shares no patients with the test folds.
Ablations. Four ablations isolate the proposal's ingredients:
- Note bias. Replace the categorical note model with a naive one that treats silence as "no problem."
- Note channel. Remove notes entirely.
- Pathology. Remove pathology from $\Lambda$ and $\mu$.
- Slide features. Replace structured pathology with whole-slide embeddings from a multiple-instance learning model.
What counts as winning. We beat the strongest baseline only if two things hold. First, the held-out domain MAE is lower than the note-covariate mixed model's, with a bootstrap confidence interval excluding zero. Second, the improvement is larger in the dropout stratum than in completers.
Why not just…?
Why not just fit mixed-effects models per domain, as READT did? This is the most serious threat. If surveys alone determine the trajectory, a mixed model with pathology covariates may match us. The masked-survey experiment below settles it. If the gain over the mixed model is concentrated in, and roughly restricted to, patients with missing surveys, the note channel is what matters. If there is no gain anywhere, the state-space machinery is not justified, and we would report the mixed model as sufficient.
Why not just add NLP note flags as covariates to that mixed model? This simpler fusion is the precise alternative to our core ingredient, and we run it as a baseline. The difference is how silence is handled. A covariate treats "no mention" as a value, while our measurement model makes the meaning of silence depend on the specialty and visit type. The mechanism behind the concern comes from the file itself: Hernandez-Boussard et al. identified explicit negative documentation in 1,035 patients, while the remaining text-evaluable patients (roughly 1,485 − 450 − 1,035 ≈ 0) apparently had neither, which makes their silent-note count hard to interpret without more detail. If the covariate version performs as well, we will conclude that the bias model is unnecessary and say so.
Why not just predict 12-month recovery from baseline pathology and covariates, with no trajectory? Static prediction is the default in surgical outcomes work. Beyond Thresholds and Meiring et al. show that static snapshots lose accuracy as time passes. Their evidence comes from physiology over hours and days, however, not from function over months. We include a static gradient-boosted classifier for 12-month non-recovery and report whether updating predictions with data from months 1–6 improves AUROC over it.
Why not just use richer slide-image features instead of the reported pathology? Automated slide analysis is accurate for diagnosis (Da Silva et al.). Nothing in the file shows that it predicts functional outcomes beyond stage, margin and grade. We therefore treat slide embeddings as an ablation, not as a claim.
Proposed Experiments
Component check: note extraction. Before any trajectory modeling, domain-wise note extraction must reach at least the PPV of 0.73 and sensitivity of 0.84 that Hernandez-Boussard et al. reported for incontinence. It is measured on the annotated snippets. A domain that falls well short is dropped from the note channel.
Component check: survey structure. We confirm the five-domain EPIC-CP structure in the linked cohort. We expect the hormonal domain to be weak, since Lourenço et al. reported α = 0.35 for vitality/hormonal, and we report it separately.
Decisive experiment: masked surveys. In patients who completed all three surveys, we artificially hide the 6-month or 12-month survey. Every method then predicts the hidden domain scores from whatever remains, with notes included where the method uses them. The metrics are:
- domain-wise MAE and calibration of predictive intervals;
- AUROC for failing to reach a prespecified continence-domain recovery threshold at 12 months.
This masking gives us ground truth for exactly the situation that real dropout creates.
Real dropout. In patients who genuinely dropped out, no ground truth is available. We compare the model's inferred states against later incidental survey returns where they exist. We also report whether the fitted dropout model says missingness depends on the latent state.
Pathology ablation. We report the change in held-out error when pathology is removed from the dynamics, and whether the effect sits in the recovery rate or in the plateau.
Hidden vulnerable patients. Following Beyond Thresholds, we examine patients whom the static model rates as likely to recover. The question is whether the dynamic model identifies, by month 3, a subgroup among them that does not recover.
Failure criteria. The core claim fails if either of two things happens:
- the note-covariate mixed model matches our model in the masked-survey experiment;
- the improvement does not grow with the amount of missing survey data.
Either outcome would tell the field that clinician notes can simply be added as covariates.
References
- New Paradigms for Patient-Centered Outcomes Research in Electronic Medical Records: An example of detecting urinary incontinence following prostatectomy
- Health-Related Quality of Life Assessment in Prostate Cancer Patient Undergoing Androgen Deprivation Therapy: Real-World Experience in the READT Study
- Independent real‐world application of a clinical‐grade automated prostate cancer detection system
- An Online Mammography Database with Biopsy Confirmed Types
- Patient-reported outcomes after surgery or radiotherapy for localised prostate cancer: a retrospective study
- Portuguese version of the Expanded Prostate Cancer Index Composite for Clinical Practice (EPIC-CP): psychometric validation and prospective application for early functional outcomes at a single institution
- Mapping the Minnesota Living with Heart Failure Questionnaire (MLHFQ) to SF-6Dv2 in Chinese patients with heart failure
- A deep state-space analysis framework for cancer patient latent state estimation and classification from EHR time-series data
- DySurv: dynamic deep learning model for survival analysis with conditional variational inference
- Beyond Thresholds: Learning Perioperative Risk from Intraoperative Physiological Trajectories
- Fine-tuning Zero-shot Large Language Models for Patient-reported Outcomes (Student Abstract)
- Optimal intensive care outcome prediction over time using machine learning