When a patient leaves the hospital, the discharge summary records what they were told to do. Sometimes it also records whether they said they would do it. When the same patient comes back, the admission notes often record what actually happened: which medications they report taking, whether they "stopped the metoprolol because of dizziness," whether they "never filled" a prescription. Each pair of consecutive admissions is therefore a small natural record of the intention-behavior gap, the well-documented distance between what people know and intend and what they actually do. This record covers thousands of patients, has hard outcomes attached, and requires no questionnaire. We propose to read these paired notes with a language model and sort each discharge instruction into one of four states. In the first, the patient was never clearly told (a knowledge failure). In the second, the patient was told and declined (no intention). In the third, the patient was told, agreed, and did not follow through (the intention-behavior gap proper). In the fourth, the instruction was carried out. We then use this phenotype as an exposure in a retrospective cohort study, meaning an analysis of past routine-care records that follows patients forward from a documented exposure to later outcomes.
The opportunity matters because both fields stall at the same point. Work on the intention-behavior gap relies on cross-sectional self-report. The dietary study in fatty liver patients found positive attitudes alongside poor practice but could not follow anyone over time (KAP in NAFLD). The "Practice Primacy" study found that habit variables carried over 63% of predictive importance, but its knowledge scale had a Cronbach's alpha of 0.313 (Practice Primacy). Retrospective cohort work has the opposite problem. It has long follow-up and hard endpoints but treats patient behavior as unmeasured noise. The statin-response analysis of ACCORD attributes its near-zero HDL predictability ($R^2 \approx 0.040$) to uncaptured lifestyle and adherence (Statin response in T2D). The 50,912-patient stroke readmission study can see behavior only at its extreme, discharge against medical advice (OR 1.485) (Stroke readmission).
Core Claim
Non-adherence is usually recorded in health records as a single flag, such as a "noncompliant" keyword or a noncompliance diagnosis code. That flag mixes three different failures that call for different interventions: not knowing, not intending, and intending but not acting. Our claim is that these three can be separated reliably from paired discharge and readmission notes. We expect them to differ in two ways. First, they should carry different prognostic information for later outcomes such as one-year mortality and further readmission. Second, they should differ in how persistent they are across a patient's admissions.
The new element is not detecting non-adherence in notes. It is decomposing non-adherence into knowledge, intention, and action components, measured at the level of individual instructions across linked admissions, and then testing in a cohort design whether that decomposition carries information a single flag does not. If the decomposition adds nothing beyond a well-built single flag, the claim is false, and the experiments below are designed to show that if it is.
Research Program
The data source is MIMIC-IV, a large de-identified hospital record database. The cohort is adult patients with at least two hospital admissions that have discharge summaries. MIMIC-IV also links to out-of-hospital death dates, so outcomes are not limited to patients who return.
For each index discharge, a clinically adapted language model extracts the actionable instructions. These cover medications started or changed, diet, follow-up appointments, and monitoring. For each instruction, the model also extracts any documented patient response, such as agreement, refusal, or expressed concern.
From the next admission's history, home-medication list, and reconciliation text, a second extraction pass assigns each instruction to one of the four states, or to "unknown." Every assignment must cite the text spans that support it. This requirement comes from the AICare user study, which found that clinicians trusted model outputs only when they could check them against visible evidence (AICare).
We know long multi-note contexts degrade language model accuracy. EHRNoteQA reports GPT-4 at 97.16% multiple-choice accuracy on discharge-summary questions, with performance falling as more admissions are packed into the context (EHRNoteQA). For that reason we pair notes one discharge-readmission link at a time rather than feeding entire patient histories.
Clinicians will annotate a gold set of instruction-state labels. To choose which cases they annotate, we use entropy-based uncertainty sampling. The cohort-construction study on MIMIC-IV found that this method beat random sampling on all 18 imbalanced tasks, whereas diversity sampling did worse than random (Active learning for EHR cohorts). Here that paper supplies method rather than motivation. It matters because "agreed but didn't act" is likely to be a rare class.
The cohort analysis treats the patient-level phenotype, measured at readmission $k+1$, as the exposure. Outcomes are measured strictly afterward: death within one year and admission $k+2$. Baseline adjustment covers demographics, comorbidity, insurance, length of stay, and discharge against advice.
Because the exposure can only be observed in patients who return, we model the probability of readmission and apply inverse-probability-of-return weights. We report results both weighted and restricted to returners.
We deliberately do not use the phenotype as a covariate in a propensity-score treatment comparison of the kind used in prostate cancer survival. Adherence after treatment lies on the causal pathway from treatment to outcome, so adjusting for it would remove part of the effect we are trying to estimate. The file raises this as a tension, and our design avoids it rather than ignoring it.
A secondary analysis tests "Practice Primacy" on hard longitudinal data. It asks whether a patient's past execution state at admission $k$ predicts execution at $k+1$ better than the clarity of the instructions they received at $k+1$.
Baselines
The naive reference point is the stroke-study approach of using only structured behavioral proxies: discharge against medical advice, insurance or self-pay status, and length of stay, entered into a multilevel logistic model (Stroke readmission). We fit it on the same MIMIC-IV cohort, with the same outcomes and the same baseline covariates. Beating it means the four-state phenotype significantly improves discrimination and calibration for one-year mortality and readmission $k+2$, assessed by bootstrap confidence intervals on the difference in AUROC (area under the ROC curve) and Brier score.
The strongest realistic competitor is a single non-adherence flag built as well as possible. It fires if the readmission notes contain a curated noncompliance lexicon ("noncompliant," "not taking," "ran out," "did not fill"), if a noncompliance diagnosis code is present, or if the same language model is simply asked "was the patient adherent?" This competitor is strong because it uses the same notes and largely the same extraction cost. It removes exactly the ingredient we claim matters, the decomposition. It gets the same model, the same prompt-tuning budget, and the same gold-set development data. Beating it requires that the four-state model improves prediction over this flag and that the separate states show hazard ratios that differ with non-overlapping intervals. No reported result was found on Lacuna for note-based adherence extraction or readmission prediction on MIMIC-IV. The nearest reported number is EHRNoteQA's 97.16% multiple-choice accuracy for GPT-4, which bounds single-note reading ability but is not a result on this task.
The ablation that isolates the key ingredient collapses the four states in two different ways. One version merges "declined" with "agreed but didn't act" (intention dropped). The other merges "never told" with "agreed but didn't act" (knowledge dropped). Comparing each with the full model shows which distinction carries the signal.
A final baseline feeds the full linked note history to an end-to-end language model that predicts the outcome directly, with no intermediate phenotype. It gets the same model and the same compute.
Why not just…?
Why not just use a noncompliance keyword or diagnosis code? This is the real threat, and we treat it as the primary baseline rather than arguing it away. Clinicians may already write "noncompliant" mostly for one failure type, in which case a flag would capture most of the signal. The collapsed-state ablations settle this directly. If merging states costs nothing in prediction or in hazard-ratio separation, we will report that the decomposition is not worth its cost.
Why not just survey patients with knowledge-attitude-practice questionnaires? Surveys measure the gap directly but only at one moment, and they rest on self-report of uncertain reliability: the alpha of 0.313 in Practice Primacy, and the self-reported adherence noise acknowledged in Statin response in T2D. They also cannot be linked retrospectively to mortality in an existing cohort. Notes are not free of bias, since clinicians may under-document refusals, so we validate against physiology as described below rather than assuming the notes are accurate.
Why not just let a large language model read all the notes and predict outcomes? It might predict well, but it would not tell anyone which failure to intervene on. EHRNoteQA also shows accuracy degrading over multi-admission contexts. It is still run as a baseline. If it beats our model on prediction, our contribution narrows to interpretability and the cohort-level hazard comparison, and we will say so.
Decisive Checks
The first check is extraction validity. We will report state-level agreement between the model and clinicians, and between clinicians themselves, on the gold set. If inter-clinician agreement on "agreed but didn't act" versus "declined" is poor, the construct is not measurable from notes and the project stops there.
The second check is a physiological test that does not depend on note-reading. For antihypertensive, diabetes, and anticoagulant instructions, admission values at $k+1$ should rank as expected: blood pressure, glucose or HbA1c where available, and INR (the standard measure of blood clotting time used to monitor anticoagulation) should be worse for "never told" and "agreed but didn't act" than for "executed." If the states do not separate on these labs, the labels are tracking how clinicians write rather than what patients did.
The third check is a negative control. Instructions that patients cannot execute themselves, such as an inpatient procedure scheduled by the hospital, should show no association between execution state and outcome.
The fourth check is the main cohort result. We test whether "agreed but didn't act" carries a mortality and readmission hazard distinct from "declined" and "never told," and whether persistence across admissions favors past practice over instruction clarity, as Practice Primacy would predict.
References
- Knowledge, attitude, and practice on dietary management among patients with non-alcoholic fatty liver disease: a cross-sectional study in Sichuan
- Practice Primacy: Revisiting the Knowledge–Action Gap in Pro-Environmental Behavior with eXplainable AI
- A Machine Learning Approach to Predicting Heterogeneous Lipid Responses to Statins in Type 2 Diabetes
- Risk factors associated with 31-day unplanned readmission in 50,912 discharged patients after stroke in China
- Augmenting Clinical Decision-Making with an Interactive and Interpretable AI Copilot: A Real-World User Study with Clinicians in Nephrology and Obstetrics
- EHRNoteQA: An LLM Benchmark for Real-World Clinical Practice Using Discharge Summaries
- Active Learning as Decision Support for EHR Cohort Construction Under Class Imbalance
- Long-Term Survival in Men With High Grade Prostate Cancer: A Comparison Between Conservative Treatment, Radiation Therapy and Radical Prostatectomy—A Propensity Scoring Approach