PaperlayerView exact revision on arXiv

Source-linked AI summary

Rolling Day-Wise Mortality Prediction in Critically Ill Patients With AKI on CRRT Utilizing Machine Pressure Waveforms

Shehan Irteza Pranto, Joanna Yang, Joshua Lambert, Stuart L. Goldstein, Lili Chan, Girish N. Nadkarni, Tiago K. Colicchio, Javier A. Neyra, Jin Chen

arXiv:2609.13524v1cs.LGcs.AI

TL;DR

Critically ill patients with AKI on CRRT face high mortality, while risk assessment relies mainly on intermittently sampled EHR data and leaves minute-level CRRT waveforms unused for patient-level mortality prediction. The paper cleans and aligns machine signals, then combines a rolling day-wise Transformer ensemble with clinical variables. The combined model achieved one-day mortality AUROC 0.766, compared with 0.717 for EHR-only features.

  • Problem

    Minute-level CRRT waveforms have not been used for patient-level mortality prediction, while EHR measurements are intermittently recorded and may miss evolving physiology.

    Source: §I · p. 1

  • Method

    The study uses a leak-safe rolling day-wise task and a two-level ensemble that late-fuses a window-reduced pressure-stream Transformer with classical models using circuit-instability and EHR features.

    Source: §I · p. 2

  • Results

    AUROC 0.766 for combined EHR and machine features exceeded the 0.717 clinical-only reference; machine-only pressure achieved AUROC 0.625 for next-day mortality.

    Source: §I · p. 2

  • Takeaways & Limitations

    SHAP attribution placed seven CRRT metrics among the combined model’s top 15 features, and signal conditioning contributed +0.029 AUROC in the reported ablation.

    Source: §V · p. 9

    External validation is required before claiming clinical utility because the CRRTnet split was patient-level rather than site-level.

    Source: §V · p. 9

Abstract

from arXiv · show

Critically ill patients with acute kidney injury (AKI) on continuous renal replacement therapy (CRRT) face high mortality, yet current risk assessment relies primarily on clinical parameters from electronic health records (EHR) and ignores minute-level circuit pressure waveforms generated by CRRT machines that track the extracorporeal circuit's interaction with the patient. Clinicians therefore cannot see deterioration as it develops. Risk is reassessed only when labs are drawn, while this continuous record is discarded because it is contaminated by shared-device records, non-physiological minutes, and sensor artifacts. To make the stream usable, we aligned machine records to charted therapy intervals to prevent cross-patient leakage, removed priming and downtime minutes, tuned denoising on a synthetic spike-injection benchmark, and masked unobserved intervals rather than imputing them. On this cleaned stream, we define a rolling day-wise task and a transformer-based stacked ensemble that late-fuses a window-reduced sequence transformer with classical models using circuit-instability features and clinical EHR variables. In a leak-safe benchmark on the multi-center CRRTnet cohort (976 patients, 4,585 treatment days), the machine-only model had the lowest standalone prognostic value (AUROC 0.625), followed by the EHR-only model (0.717). Integrating EHR and machine streams reached a one-day mortality AUROC of 0.766. SHAP attribution showed that circuit-instability descriptors raised the machine share of the top 15 combined-model features from 3 to 7 (20.0% to 46.7%), highlighting filter pressure, transmembrane pressure (TMP), and access-to-return difference (ARD). To our knowledge, this is the first patient-level mortality prediction incorporating CRRT machine data, turning a discarded bedside stream into a continuous risk signal.

I. INTRODUCTION

The paper addresses the unexplored use of minute-level CRRT waveforms for patient-level mortality prediction, alongside limitations of intermittently sampled EHR data. It proposes a leak-safe rolling day-wise framework and reports improved discrimination when machine signals are combined with EHR features.

I. INTRODUCTION

  • Minute-level CRRT waveforms had not been used for patient-level mortality prediction in a rolling, day-wise framework, while EHR-based approaches relied on intermittently recorded measurements.The introduction notes that high-resolution machine data had instead been used for circuit-level tasks such as clotting or circuit-lifespan prediction.Source: §I · pp. 1–2
  • Shared-device records, non-physiological minutes, sensor artifacts, and interpolation across unobserved intervals can make extracted waveforms unreliable and inflate apparent performance.The paper argues that establishing the extracted waveform is real is necessary before assessing its prognostic value.Source: §I · p. 2
  • AUROC 0.766 for the combined model exceeded 0.717 for the clinical-only reference, while machine-only pressure predicted next-day mortality with AUROC 0.625.The authors report this as evidence that circuit-pressure signals carry patient-level prognostic information beyond routine EHR features.Source: §I · p. 2
  • The proposed framework scores each treatment day from its minute-level CRRT signal using a causal rolling task with an excluded gap and a two-level ensemble late-fusing a Transformer with classical models.The ensemble also uses engineered circuit-instability features and clinical EHR variables, with preprocessing and model-selection decisions restricted to training or validation data.Source: §I · p. 2, Figure 1 · p. 2

II. COHORT AND PREDICTION TASK

The study evaluates linked EHR and CRRT machine data from a multicenter cohort, with treatment-day mortality predicted in a rolling, causal setup. Its signal-processing pipeline aligns machine records to therapy intervals, removes non-physiological minutes, and denoises pressure signals.

A. Cohort

  • The evaluated CRRTnet cohort included 976 patients and 4,585 treatment days, restricted to the first seven treatment days and split at the patient level into training, validation, and held-out test sets.The registry covers adults with AKI on CRRT and links clinical EHR data with machine data.Source: A. Cohort · p. 2

B. Rolling day-wise labeling

  • Each treatment day’s CRRT stream and available EHR data form the input, followed by an excluded one-day intervention gap and a one-day mortality prediction horizon.The window advances one day at a time; baseline covariates and forward-filled labs are used without look-ahead.Source: B. Rolling day-wise labeling · p. 2

A. Signal acquisition and cleaning

  • Pressure spikes are detected by deviations from a window mean relative to local standard deviation, then replaced using neighboring valid samples; denoising settings are selected on a synthetic spike-injection benchmark.The benchmark uses 5% corruption and evaluates spike-localization F1 and reconstruction error.Source: Figure 2 · p. 3, A. Signal acquisition and cleaning · p. 3
  • Machine timestamps are restricted to charted therapy intervals, with a 12-hour buffer after the stop time; this alignment removed 13.5% of raw machine data.The buffer accommodates delayed charting without admitting a neighboring patient’s session.Source: A. Signal acquisition and cleaning · pp. 2–3
  • Priming minutes identified by zero run-time and downtime minutes identified by zero effluent increments are excluded, removing a further 1.8% of minutes.These intervals reflect device operating states rather than patient states and can bias daily distributional summaries.Source: A. Signal acquisition and cleaning · p. 3

B. Data Structuring and Derived Physiology

The section formats denoised pressure streams for day-wise modeling, preserves unsupported intervals as missing, and derives pressure gradients and engineered features to represent circuit physiology.

B. Data Structuring and Derived Physiology

  • Long gaps and out-of-bounds minutes remain masked, while only short dropouts are linearly interpolated.The distinction avoids fabricating unobserved physiological trends while retaining a fixed-length day representation.Source: B. Data Structuring and Derived… · p. 3
  • TMP, PFD, and ARD represent membrane, hemofilter, and access-to-return pressure gradients; ARD is used only by the engineered-feature branch.The deep model receives access, filter, effluent, return, TMP, and PFD channels together with a validity mask.Source: B. Data Structuring and Derived… · p. 3
  • The full 200-feature machine set is reduced by mRMR, with a 40-feature machine budget in the deployed model.The machine and clinical modalities are selected separately before modeling.Source: Table II · p. 4, B. Data Structuring and Derived… · p. 4
  • The engineered branch computes 200 features: 133 base statistics and 67 enriched descriptors, including spectral, entropy, excursion, slope, and cross-channel measures.Features are computed over valid minutes only.Source: B. Data Structuring and Derived… · p. 4

C. Stacked ensemble of a Transformer-fusion network and classical learners

The section describes a stacked ensemble and three input representations: a window-reduced machine sequence, a clinical vector, and a grouped tabular vector.

C. Stacked ensemble of a Transformer-fusion network and classical learners

  • A logistic-regression meta-learner combines the Transformer-fusion network and classical tabular learners into the final estimate Pstack.The section’s input representations supply the sequence and tabular branches of this stack.Source: C. Stacked ensemble of a… · p. 4

1) Input representations (Fig. 3A):

  • The machine sequence comprises 144 non-overlapping 10-minute windows, each summarized into a 30-dimensional token; windows without valid minutes are masked.Each token uses five statistics: mean, minimum, maximum, standard deviation, and least-squares slope.Source: 1) Input representations (Fig. 3A) · p. 4
  • The deep network’s clinical view uses the top 60 mRMR-selected patient features, while classical learners receive 40 machine features concatenated with 60 clinical features.The grouped feature budget is selected on validation.Source: 1) Input representations (Fig. 3A) · p. 4

2) Transformer-Fusion Deep Network (PDL) (Fig. 3B):

The deep network separately encodes the masked machine sequence and clinical vector, then combines their embeddings to produce a mortality probability.

2) Transformer-Fusion Deep Network (PDL) (Fig. 3B):

  • The Transformer encodes the window sequence with positional information and masks invalid or gap windows so only observed signal contributes.The selected configuration uses one encoder layer, four attention heads, hidden width 64, feed-forward width 128, and dropout 0.4.Source: 2) Transformer-Fusion Deep Network… · p. 4
  • A small MLP maps the reduced clinical vector to an embedding, while the Transformer’s summary-token state supplies the machine embedding.The clinical block includes batch normalization, ReLU, and dropout.Source: 2) Transformer-Fusion Deep Network… · p. 4
  • The fusion head concatenates the two embeddings and applies a two-layer head to produce the deep-network mortality probability PDL.The configuration was selected by validation AUROC from 12 candidate settings.Source: 2) Transformer-Fusion Deep Network… · pp. 4–5, Figure 3 · p. 5

3) Classical Base Learners (P1, . . . , PK) (Fig. 3C):

The classical branch trains several tabular learner families and passes the validation-ranked top three into the ensemble stack.

3) Classical Base Learners (P1, . . . , PK) (Fig. 3C):

4) Stacking Meta-Learner (Pstack) (Fig. 3D):

A logistic-regression meta-learner combines the deep-network and top-three classical-model probabilities, with stacking and model-selection choices kept off the test split.

4) Stacking Meta-Learner (Pstack) (Fig. 3D):

  • The stack combines the deep-network probability and top-three classical probabilities as meta-features in a logistic-regression meta-learner.Its weights are fit on validation probabilities and labels; the binary decision threshold is also selected on validation by maximizing F1.Source: 4) Stacking Meta-Learner (Pstack)… · p. 5

D. Training, Selection, and Leakage Control

  • Class-weighted binary cross-entropy and weighted sampling address the roughly 1:8 day-level class ratio during deep-network training.Training uses early stopping on validation AUROC and a best-of-five-seeds checkpoint.Source: D. Training, Selection, and Leakage… · p. 5
  • The train split fits preprocessing, feature selection, neural weights, and classical-model tuning; validation fits stacking and selection choices, while test is reserved for final performance.The test split is untouched during fitting.Source: D. Training, Selection, and Leakage… · pp. 5–6

IV. EXPERIMENTS & RESULTS

The experiments validate a fixed denoising setting and compare clinical-only, machine-only, and combined pipelines. The combined stack achieves the highest reported one-day mortality AUROC, while the waveform alone also carries prognostic signal.

IV. EXPERIMENTS & RESULTS

  • The evaluation uses day-level test AUROC as its primary metric and also scores full, 1:1, and 1:2 prevalence protocols.The deep sequence baselines span convolutional, recurrent, and attention-based families, with a clinical MLP reference.Source: §IV · p. 6, B. Three-Pipeline Benchmark · p. 6

A. Denoising validation

  • L = 40, κ = 2 with mean smoothing over η = 2 neighbors maximized detection F1 (0.850) and minimized reconstruction error (RMSE 9.14).The operating point was selected using a synthetic spike-injection benchmark and held fixed across the reported pipelines.Source: A. Denoising validation · p. 6, Figure 4 · p. 6

B. Three-Pipeline Benchmark

  • The day-level test AUROC varied narrowly from 0.719 to 0.766 across feature budgets, with no trend in either budget axis.The validation-selected common budget was used for all classical and deep models, avoiding feature-count confounding in their comparisons.Source: B. Three-Pipeline Benchmark · p. 6
  • 0.766 AUROC for the combined stack exceeds the clinical-only reference (0.717), while machine-only reaches 0.625 and the best combined classical learner reaches 0.728.The combined stack also exceeds the best deep fusion baseline, Attention-CNN1D (0.682); the benchmark tables cover deep-learning and classical models across the three pipelines.Source: B. Three-Pipeline Benchmark · p. 6, Table III · p. 7, Table IV · p. 7

C. Ablation Study

The ablations show that both signal conditioning and the window-reduced sequence representation contribute to the deployed model's performance. Enriched circuit-instability features shift attribution toward machine variables without measurable AUROC cost.

Preprocessing

  • 0.737 AUROC on raw signal rose to 0.766 (+0.029) with cumulative cleaning, with peak removal the largest single step (+0.019).Each individual increment was smaller than the ±0.06 test confidence interval, so the evidence supports the cumulative pipeline effect rather than crediting an isolated module.Source: C. Ablation Study · p. 7, Table V · p. 8
  • The enriched feature set raised the machine share of the top-15 SHAP features from 20.0% to 46.7% without measurable AUROC cost.Its contribution is attribution rather than discrimination.Source: C. Ablation Study · p. 7, Table V · p. 8

Sequence encoder

  • 0.766 AUROC for the windowed Transformer exceeded Patch-Transformer (0.732), CNN1D (0.729), and RNN (0.713).The 0.034 gap over Patch-Transformer attributes the gain to the window-reduced representation rather than self-attention alone.Source: C. Ablation Study · pp. 7–8, Table V · p. 8

D. Training Dynamics

Training losses stayed close across epochs, while SHAP analysis found enriched circuit descriptors among top mortality predictors alongside systemic markers.

D. Training Dynamics

  • Both training and validation losses decreased, with a 0.03–0.05 gap that widened modestly after epoch 17; the best-validation checkpoint was selected per seed.The authors describe this pattern as indicating the fusion core is well regularized.Source: D. Training Dynamics · p. 8

E. Interpretability and Clinical Significance

  • Elevated INR, low platelets, lactate, bicarbonate, pH, albumin, and phosphate reproduced a clinically relevant systemic-decompensation risk profile.The authors present recovery of this established clinical profile as a validity check before adding the machine signal.Source: E. Interpretability and Clinical… · p. 8
  • Filter-pressure permutation entropy ranked second overall; elevated TMP, pressure excursions, and ARD skewness also entered the top 15.The paper links erratic filter pressure and sustained TMP elevation to progressive micro-clotting and membrane fouling, and access-pressure instability to vascular-access dysfunction.Source: E. Interpretability and Clinical… · p. 8
  • Circuit descriptors update every minute, compared with labs drawn once or twice daily, and every fusion model beats its clinical-only counterpart.The waveform alone was weakly prognostic (AUROC 0.625); the paper frames it as complementing, not replacing, laboratory risk assessment.Source: E. Interpretability and Clinical… · p. 8
  • 3 to 7 of the top 15 features were CRRT machine descriptors after enrichment (20.0% to 46.7%).The held-out-test SHAP analysis used the stack’s Random-Forest tabular learner; the beeswarm interleaves machine descriptors with systemic markers.Source: E. Interpretability and Clinical… · p. 8, Table VI · p. 9, Figure 7 · p. 9

V. DISCUSSION AND CONCLUSION

The study finds that CRRT waveforms add prognostic information to routine EHR features, but external validation and more informative labels are needed before bedside use.

V. DISCUSSION AND CONCLUSION

  • 0.766 AUROC was reached by fusing waveforms with EHR features, above the best clinical-only model (0.717); machine-only AUROC was 0.625.The study reports that seven CRRT metrics appeared among the combined model’s top 15 features.Source: §V · p. 9
  • +0.029 AUROC came from conditioning the raw stream, making acquisition and cleaning part of the reported finding.The gain was on the same order as the encoder choice.Source: §V · p. 9
  • External validation is required before claiming clinical utility because the CRRTnet split was patient-level, not site-level.The authors identify an independently collected University of Alabama at Birmingham cohort as a transportability test.Source: §V · p. 9
  • About 6.4% of treatment days were positive (62 in the test split), widening confidence intervals and limiting fine comparisons between configurations.The day-level label assigns one outcome to all 1,440 minutes, although the informative intervals within a day are unknown.Source: §V · p. 9
  • The system is not ready for standalone decision support, and reliability within specific patient phenotypes remains to be established.The authors identify per-phenotype reliability as a necessary step toward trustworthy decision support.Source: §V · p. 9
Loading 2609.13524v1…