Hospital risk models built from administrative records predict readmission and death well using only billing codes, procedure counts and utilization. They treat each hospital stay as a series of transactions. A separate literature finds that how patients are treated during an encounter changes what they do afterward. In one study, patients who felt pressured during a screening consultation were about twice as likely to drop out of the program over four years. Clinicians also routinely write interpersonal content into their notes: the patient's own words, their goals, their social context. This proposal asks whether that documented interpersonal content, extracted automatically from routine discharge notes, predicts what happens after discharge beyond what billing codes already capture. It also asks whether any such signal survives once we rule out the boring explanation that some clinicians simply write longer, more careful notes.
The two-sentence version is this. We build a text classifier that tags patient-centered-care content in hospital discharge notes, then run a retrospective cohort study (a study that follows patients forward through existing records from a historical start date) to test whether these tags predict 30-day readmission after controlling for administrative codes, note volume and stable patient traits. We then test, with a small set of pre-specified subgroups, whether both the documentation rates and the prognostic value differ by race, age and insurance. This matters because uneven documentation could quietly feed bias into clinical prediction models.
Core Claim
The claim is narrow and falsifiable. Suppose a patient's discharge documentation contains patient-centered-care content: patient voice, patient goals, social context, shared decision-making, provider support for goals, or mention of a caregiver. We claim this content carries information about 30-day readmission that is not carried by three other sources. The first is structured administrative codes. The second is generic measures of how much was written. The third is a general-purpose embedding of the same note text.
A secondary claim is that these concepts are documented at different rates across racial and insurance subgroups, even at matched clinical severity. Adding them to a codes-only model therefore changes how well that model is calibrated within subgroups.
If the incremental signal disappears once note volume is controlled, the honest conclusion is that "interpersonal documentation" is a proxy for documentation effort. That negative result would also be worth publishing.
What Exists and What Is Missing
The measurement idea comes from Clinician documentation of patient centered care in the electronic health record. That study hand-coded seven patient-centered-care concepts in 477 Veterans Affairs notes and found patient voice in 67.7% of notes. It also found that sites with higher patient-satisfaction scores documented patient goals more often. However, satisfaction was measured only at the site level, and no patient was followed forward. We reuse its seven-concept scheme but link the concepts to individual patients' later outcomes.
The extraction method comes from Automated Scoring of Communication Skills in Physician-Patient Interaction. That work scored empathy and information-giving with DeBERTa-large and reached a mean binary F1 of 0.94 on transcribed simulated vignettes. Real discharge notes are messier and more abbreviated than transcripts, so we expect lower accuracy and will report it rather than assume it.
The motivating evidence that interpersonal experience changes later behavior comes from two survey-based studies. Regret on Choice of Colorectal Cancer Screening Modality found that regret after the consultation raised four-year non-compliance (odds ratio 2.189). How does people-centered integrated care ... promote the continuity of healthcare for internal migrants found that provider respect strengthened the effect of care coordination only in economically developed regions. Both relied on bespoke surveys. Neither touched routine records.
On the administrative side, two studies show strong prediction from codes alone. Death vs. Data Science deliberately bypassed note text and still reached AUC 0.79 from claims. Application of Machine Learning Techniques ... After Stroke Using Electronic Administrative Records showed that 1,303 administrative features outperformed a clinical registry model. These results constrain us: the codes-only model is a strong opponent, not a straw man.
Finally, An Extensive Data Processing Pipeline for MIMIC-IV supplies two things. It supplies the cohort extraction and readmission targets. It also documents a stark subgroup disparity: an LSTM predicting non-ICU mortality had a positive prediction rate of 0.0102 for White patients and 0.0024 for African American patients. Our subgroup analysis asks whether documentation disparities are one route by which such gaps arise.
Research Program
Measurement. We will annotate about 1,500 discharge-note passages from MIMIC-IV for Lin et al.'s seven concepts. Two annotators will label each passage, and we will report inter-annotator agreement. We will then fine-tune a DeBERTa-large tagger following the ACTA recipe, including synthetic augmentation for rare concepts. Any concept with held-out F1 below 0.8 is dropped from the main analysis rather than used noisily. Lin et al. report that shared decision-making was rarely documented, so we expect it may fail this bar.
Volume controls. Alongside the concept tags, we compute documentation-volume features for every note. These are total token count, number of sections, and counts of non-interpersonal clinical entities such as medications and procedures. They let us separate "talked about the patient as a person" from "wrote a lot."
Cohort. The cohort is adult MIMIC-IV hospital admissions discharged alive, extracted with the Purushotham et al. pipeline using temporal splits to avoid leakage. The index date is discharge. The primary outcome is unplanned 30-day readmission. A secondary behavioral outcome is subsequent discharge against medical advice. If documented interpersonal content tracks the quality of the relationship, it should relate to this outcome.
Analysis. There are three layers. The first is predictive: we measure the incremental area under the ROC curve (AUROC) of the concept tags over nested baselines. The second is within-patient. Among patients with several admissions, a conditional logistic model with a fixed effect for each patient asks whether changes in documented concepts between that patient's admissions track changes in readmission. Because each patient serves as their own control, stable traits such as personality, chronic severity and social circumstances cancel out. The third is subgroup analysis. We pre-specify four strata: race, age 80 or older, Medicaid insurance, and limited English proficiency. Heterogeneity is tested with interaction terms and corrected with the Holm procedure, which controls the false-positive rate across multiple tests.
We will not search for subgroups. The multiplicity and power problems with exploratory stratification are the main known weakness of that literature.
Baselines
The naive reference is a logistic regression on age, sex, admission type and a comorbidity count. This is what a clinician-designed score would roughly capture. It tells us how much any added model buys.
The strongest structured competitor is gradient-boosted trees on the full set of grouped diagnosis codes, procedure codes, medications, prior utilization and length of stay. It is built with the Purushotham et al. pipeline, which is the closest available analogue to the claims-only models of Ahmad et al. and Talman et al. We found no reported readmission result on Lacuna for this pipeline on MIMIC-IV. The nearest reported number in the file is for a different task: ICU mortality AUROC of 0.87 for a hybrid LSTM versus 0.67 for logistic regression in a chronic kidney disease cohort. So we will establish the readmission number ourselves on identical splits.
The strongest text competitor is the same tree model plus a PubMedBERT embedding of the full discharge note. This follows Revisiting Clinical Outcome Prediction for MIMIC-IV, which reports ED diagnosis AUROC of 86.86% with PubMedBERT. This baseline tests whether our specific concepts add anything that generic text features do not.
The key ablation is codes plus volume features, with and without the concept tags. All models share splits, receive the same hyperparameter search budget (equal numbers of tuning trials on the validation fold), and are scored on the same held-out later period.
We set the success threshold before seeing results. The concept tags must improve AUROC over codes plus volume by at least 0.01, with a bootstrap confidence interval excluding zero. In addition, adding concept tags to the PubMedBERT model must not be fully redundant: the gain must be non-zero, or the concept tags alone must match the embedding's gain with seven interpretable features instead of hundreds.
Why not just…?
Why not just use billing codes, since they already predict well? Codes may be enough, and our nested design is built to show that if it is true. The evidence for a gap is indirect. Codes act as crude proxies for function, as with the incontinence code in Talman et al. The interpersonal literature points to behavioral drivers of dropout that codes cannot see. If the concept tags add nothing over codes plus volume, we will report that administrative data already absorbs this signal.
Why not just use note length? This is the most serious threat. Longer notes plausibly reflect more careful clinicians, more complex patients, or both, and that alone could produce an apparent "interpersonal" effect. We take it seriously in three ways. Volume features are included in every comparison model. The within-patient model compares the same patient's admissions. We also run a negative-control check: the concept tags at the index admission should not "predict" the number of admissions in the year before it. If they do, the tags are capturing illness burden, and we will downgrade any causal reading.
Why not just use a general text embedding? A generic embedding may outperform seven tags on raw AUROC, and we do not claim otherwise. The claim is about a named, auditable quantity. Only explicit concept tags let us ask whether patient voice is recorded less often for Medicaid patients at matched severity, and a dense embedding cannot answer that disparity question. If the embedding captures all of the tags' signal and the tags show no subgroup disparity, the measurement contribution reduces to interpretability, and we will say so.
Proposed Experiments and Decisive Checks
The first experiment validates measurement. We report per-concept F1 and annotator agreement. We also report convergent validity: the association between concept density and subsequent discharge against medical advice, which an interpersonal construct should predict.
The second experiment is the nested prediction comparison above. It is reported overall and within each pre-specified stratum, with calibration slopes per stratum.
The third experiment is the within-patient fixed-effects estimate, reported as an odds ratio per standard deviation of concept density, alongside the negative-control result.
The fourth experiment is the documentation-disparity audit. We regress concept presence on subgroup membership, adjusting for the codes-model risk score and note volume. We then test whether adding the tags narrows or widens the subgroup gaps in positive prediction rate of the kind Purushotham et al. reported.
The decisive check is whether the concept-tag gain survives both the volume controls and the within-patient design. If it survives both, routine notes contain a scalable, individual-level interpersonal signal that the survey-based literature could never link to outcomes. If it does not, the field learns that "documented patient-centeredness" in records is mostly documentation effort. Any model using it as a quality metric should then be treated with suspicion.
References
- Clinician documentation of patient centered care in the electronic health record
- Automated Scoring of Communication Skills in Physician-Patient Interaction: Balancing Performance and Scalability
- Regret on Choice of Colorectal Cancer Screening Modality Was Associated with Poorer Screening Compliance: A 4-Year Prospective Cohort Study
- How does people-centered integrated care in medical alliance in China promote the continuity of healthcare for internal migrants: The moderating role of respect
- Death vs. Data Science: Predicting End of Life
- Application of Machine Learning Techniques to Identify Data Reliability and Factors Affecting Outcome After Stroke Using Electronic Administrative Records
- An Extensive Data Processing Pipeline for MIMIC-IV
- Revisiting Clinical Outcome Prediction for MIMIC-IV