COPD 30-day readmission: adding ECG, chest X-ray and clinical notes to the structured record brings nothing detectable once temporal leakage is fixed
Three researchers test, in COPD patients from the MIMIC-IV database, whether adding an electrocardiogram, a chest X-ray and clinical notes to the structured record helps predict 30-day readmission. After fixing three timeline problems that affected up to 58% of patients depending on the modality, they find no detectable gain in discrimination or calibration, and the structured record alone remains the best. The result is useful as a methodological safeguard, but it comes from a preprint on 1,279 patients from one hospital, with a test set of only 192: it shows no evidence of a gain, not evidence of no gain.
The context
COPD (chronic obstructive pulmonary disease) is a chronic lung disease whose hospitalisations often recur. Predicting who will return within 30 days of discharge could, in theory, target closer follow-up. Structured record variables (blood tests, vital signs, diagnosis history) have long supplied most of the signal.
A popular hypothesis in medical AI holds that adding images, physiological signals or text mechanically improves prediction. The authors, Marco Pozza, Nicolò Navarin and Silvia Gabrielli (University of Padova and Fondazione Bruno Kessler, Trento), note that this is rarely tested against noise, confounding or information leakage introduced when the cohort is built. The text is a preprint posted on medRxiv on 15 October 2026, not peer-reviewed. No competing interests are declared; we did not find a funding statement in the text we consulted.
The method
The data. MIMIC-IV v3.1, a public database of hospital records from Beth Israel Deaconess Medical Center (Boston), with its extensions for notes, chest X-rays (MIMIC-CXR-JPG) and ECGs. Patients have COPD and a non-ICU admission; the outcome is all-cause 30-day readmission, not only COPD-related. A stratified sample of 2,000 patients (1,000 per outcome) yields 1,279 patients after exclusions, split by patient into 895 for training, 192 for validation and 192 for testing. The natural readmission prevalence is 18.1% (2,506 of 13,818 eligible admissions); in the enriched cohort it reaches 54.7% at test.
The leakage audit. Temporal leakage occurs when information from after the moment of prediction slips into the model inputs. The authors checked that each examination is linked to the index admission window (between admission and discharge). They corrected three alignment problems, which affected 5–9% of patients for notes, 29–43% for chest X-rays and 53–58% for ECGs.
The models. Each modality goes through a frozen encoder, meaning pretrained and left unchanged: ECG-JEPA for the ECG, Bio_ClinicalBERT for notes, the BiomedCLIP image encoder for the X-ray. The structured record goes through a small network trained for this task. Three fusion architectures are compared: plain concatenation of the vectors, scalar gating (one learned weight per modality, absent modalities masked) and a residual fusion anchored on the structured record. Coverage is very uneven: ECG 43–46%, notes 73–79%, X-rays 13–16%, structured record 100%.
The evaluation. AUC is the probability that the model gives a higher score to a readmitted patient than to a non-readmitted one. AUC differences are tested by paired patient-level bootstrap (10,000 resamples), five seeds per architecture. Calibration, i.e. agreement between announced probability and observed frequency, is assessed with the Brier score (mean squared error of probabilities, lower is better), the calibration slope and reliability diagrams, before and after a logistic recalibration.
The results
ML metrics. Used alone, the unstructured modalities are close to chance, with an AUC of 0.46–0.53, against about 0.63 for the structured record. In fusion, no architecture exceeds the structured record alone: AUC differences (structured minus fusion) are +0.0391 [−0.0459; +0.1256] for gating, +0.0121 [−0.0694; +0.0936] for concatenation and +0.0319 [−0.0546; +0.1186] for residual fusion, all with intervals containing zero. When auxiliary modalities are progressively removed, AUC improves rather than drops.
Raw calibration seemed to favour fusion (Brier about 0.24 versus 0.32), but the structured record alone was simply overconfident, with a calibration slope of 0.12 (1 would be ideal). After recalibration, Brier is 0.239 for the structured record and 0.240–0.244 for fusion, with differences whose intervals include zero. Brier skill scores, 0.016–0.036 against a constant predictor (0.2478), are minimal.
Clinical translation. At the natural prevalence of 18.1%, out of 1,000 patients discharged after COPD, about 181 would be readmitted within 30 days. An AUC of 0.63 is weak discrimination: the score could not reliably single out those 181, and the study reports neither sensitivity nor predictive value at an operating threshold. Above all, since the cohort is enriched to about 50%, the recalibrated probabilities do not apply to the general COPD population.
What is good
The leakage audit is concrete. Rather than assuming the links between tables are right, the authors measure the problem (29–58% of patients affected for X-rays and ECGs) and fix it. This is exactly the check many multimodal MIMIC studies lack.
Calibration is treated seriously. Showing that fusion's Brier advantage was only an artefact of the comparator's overconfidence, and that it disappears after simple recalibration, is a test few papers apply.
A negative result is published with its intervals. Differences are reported with 95% confidence intervals, limitations are listed in detail, and the authors acknowledge that the stratified sampling step could not be rerun because intermediate outputs were not archived.
What is less good
A test set of 192 patients cannot establish absence of effect. The confidence intervals of the AUC differences run from about −0.07 to +0.12: compatible with zero, but also with a non-negligible gain or loss. For X-rays, only 25 test patients, with a training loss collapsing to 0.006 (memorisation): the analysis is inconclusive, as the authors admit. Failure mode: insufficient statistical power.
The comparator is not treated equally. The structured-record encoder is trained in a supervised way on the outcome, whereas the ECG, text and image encoders are frozen and self-supervised. The structured record therefore starts with an advantage, and the residual fusion received no hyperparameter search of its own. Failure mode: biased comparator.
One hospital, a fuzzy target, an enriched cohort. The data come from a single centre, with no external validation. All-cause readmission mixes causes unrelated to COPD, which may dilute a signal specific to certain modalities. The cohort enriched to about 50% complicates interpretation of metrics and probabilities. The code is deposited on Zenodo with restricted access, on request for accredited PhysioNet users: reproduction is not immediate. Failure modes: population bias, misleading metric.
What it changes
For the research community, the paper argues for two simple practices: check that each modality is linked to the admission window before modelling, and report calibration after recalibration. It also suggests testing an exacerbation-specific target, cross-modal attention, and accounting for when each modality is measured relative to the moment of prediction.
For clinicians, nothing changes: the study validates no tool and is not proof that these examinations are useless for other decisions. It reminds us that, for this target, information close to discharge matters more than one-off snapshots taken earlier during the stay.
For patients and the public, the message is to be wary of the idea that "more data" means "better prediction". Adding modalities costs collection effort and complexity; what they bring must be demonstrated.
Further reading
The paper: Auxiliary modalities add no detectable discrimination or calibration over structured EHR for 30-day COPD readmission, medRxiv, DOI 10.64898/2026.10.14.26365503. Data: MIMIC-IV, MIMIC-IV-Note, MIMIC-CXR-JPG and MIMIC-IV-ECG on PhysioNet (credentialed access). Code: Zenodo, DOI 10.5281/zenodo.23096348 (restricted access on request).
Editorial transparency: this analysis was produced with the help of AI tools from the preprint text; translations are automatic.