Source-linked AI summary
Predicting Future Organ Dysfunction in ICU Patients Using Temporal Convolutional Networks on MIMIC-IV Data
Razan Albouq, Asra Aslam
TL;DR
The paper addresses whether ICU data can forecast continuous future SOFA dysfunction rather than use SOFA only for mortality prediction. It trains a residual TCN on MIMIC-IV time series and analyzes organ-system contributions and trajectory clusters, finding meaningful predictive signal but strong dependence on recent observations and important selection-related limitations.
Problem
Prior machine-learning work largely treated SOFA as an input to binary mortality prediction, leaving direct continuous trajectory forecasting and organ-system and phenotype analysis open.
Method
A residual TCN predicts next-day SOFA from three-day ICU windows, supplemented by SHAP analysis, organ-system contribution analysis, and unsupervised trajectory clustering.
Results
The TCN extracted meaningful predictive signal, while cardiovascular dysfunction most strongly discriminated severity and deterioration and clustering identified improving and persistently severe phenotypes.
Takeaways & Limitations
TCNs can predict short-term SOFA trajectories, but the model primarily anchors predictions to the latest observation rather than exploiting temporal dependencies.
Takeaways & Limitations
Complete-case selection constrained performance at low-severity levels and in deterioration detection.
Abstract
from arXiv · showhide
Predicting future organ dysfunction in Intensive Care Unit (ICU) patients is critical for early clinical intervention, yet existing machine learning approaches have largely treated the Sequential Organ Failure Assessment (SOFA) score as an input to binary mortality prediction rather than as a continuous clinical outcome in its own right. We investigate the extent to which a Temporal Convolutional Net work (TCN) can predict next-day SOFA scores from multivariate ICU time-series data extracted from MIMIC-IV, characterise the relative contribution of each organ system to total SOFA variance and deterioration, and identify distinct trajectory patterns across ICU stays. A residual TCN trained on three-day sliding windows achieved a five-fold cross-validation R2 of 0.740 +- 0.013 and MAE of 1.431 +- 0.022, outperforming a naive persistence baseline on RMSE and R2. SHAP interpretability analysis revealed that the model functions primarily as a severity-anchoring mechanism rather than a true sequence model, with predictions dominated almost entirely by the most recent observation day. Cardiovascular dysfunction emerged as the strongest discriminator of both cross-sectional severity and acute deterioration, and unsupervised trajectory clustering identified two clinically meaningful phenotypes, an improving group (58.9%) and a persistently severe group (41.1%), differentiated by cardiovascular, hepatic, coagulation, and renal involvement. We conclude that TCNs can extract meaningful predictive signal from ICU physiological data, but that short input windows and complete-case selection bias currently limit their clinical utility, motivating future work on longer input horizons, alternative missing-data strategies, and external validation.
1 Introduction
This paper addresses whether ICU time-series data can forecast continuous next-day SOFA scores, identify the organ systems most associated with severity and deterioration, and reveal distinct SOFA trajectories. It uses a residual TCN and complementary component and clustering analyses to study these questions.
- ICU monitoring must support earlier recognition of organ failure amid pressure from rising demand, limited resources, and workforce shortages.
- SOFA summarizes six organ systems with subscores from 0 to 4, producing a total score from 0 to 24 that correlates with mortality risk.
- Prior work generally used SOFA as an input to binary mortality models rather than forecasting organ dysfunction as a continuous outcome.
- The study asks how accurately a TCN predicts next-day SOFA, which organ systems drive variance and deterioration, and which trajectory phenotypes exist.
- A residual TCN predicts continuous SOFA trajectories, component analyses quantify organ-system contributions, and unsupervised clustering identifies ICU stay subgroups.
2 Related Work
Related work has mainly predicted mortality or discrete SOFA changes, while this study positions TCNs as a way to model continuous SOFA trajectories despite irregular ICU measurements and missing data.
- Earlier studies used dynamic or time-incorporated SOFA information primarily to predict post-discharge or in-ICU mortality.
- Prior SOFA prediction included binary five-day worsening or improvement and continuous regression, but used non-sequential architectures and narrower settings.
- TCNs use dilated causal convolutions to capture long-range dependencies while supporting parallel sequence processing.
- Irregular sampling and common missing values make recurrent modeling difficult in ICU electronic health record data.
- These properties motivate TCNs for continuous SOFA trajectory prediction across organ systems.
3 Methods
The methods construct daily complete-case SOFA sequences from MIMIC-IV, train a residual TCN on three-day windows for next-day prediction, and apply SHAP, component analysis, and trajectory clustering.
- Data preprocessing: MIMIC-IV data from over 76,000 ICU admissions supplied vital signs, laboratory values, and vasopressor rates needed for all six SOFA components.
- Data preprocessing: Measurements were aligned to ICU days, cleaned against physiological ranges, and aggregated using daily worst values for each variable.
- Data preprocessing: PaO2 readings were paired with nearest same-stay FiO2 measurements, retaining the lowest daily P/F ratio for respiratory scoring.
- Data preprocessing: Complete-case analysis retained patient-days with all six SOFA components observed despite heterogeneous clinical measurement frequencies.
- Model inputs: Three consecutive ICU days formed each input window, with the immediately following SOFA score as target; ICU stays were split before sequence construction.
- TCN architecture: The residual TCN used four dilated blocks with dilation rates 1, 2, 4, and 8, causal convolutions, normalization, ReLU, dropout, and skip connections.
- Interpretability and analysis: SHAP values quantified physiological feature influence, while component analysis compared variance, correlations, severity bands, and stable versus deteriorating days.
- Trajectory clustering: K-means clustering compared interpolated seven-point SOFA trajectories, with cluster number selected using silhouette scores across k values from 2 to 7.
4 Performance Evaluation
The residual TCN showed stable overall regression performance but struggled with low-severity cases, abrupt trajectory changes, and deterioration detection. Interpretability analyses indicate predictions were anchored mainly to recent observations, while cardiovascular dysfunction was the strongest discriminator of SOFA severity and deterioration.
- Regression performance: 1.431 ± 0.022 MAE and 0.740 ± 0.013 R2 were achieved across five ICU-stay-level cross-validation folds.Performance was stable across folds, with R2 ranging from 0.721 to 0.759.
- Regression performance: The TCN improved holdout RMSE by 0.027 and R2 by 0.008 versus persistence, but its MAE was higher.The TCN reduced larger errors while sacrificing some average absolute accuracy.
- Severity-stratified performance: 0.134 was the only positive R2 across severity bands, occurring in the Severe group; Moderate and Low bands yielded −0.363 and −6.256.Complete-case selection concentrated the training distribution around moderate-to-severe scores.
- Deterioration detection: The deterioration detector identified 19 of 733 deteriorating sequences, with recall 0.03 and F1-score 0.05 for the deteriorating class.It correctly identified 4,147 of 4,181 stable sequences, reflecting substantial class imbalance.
- Trajectory behavior: Predictions tracked gradual trajectories but underestimated or missed abrupt changes, including deterioration peaks by approximately 2–3 SOFA points.This smoothing bias was consistent with the model’s dependence on the most recent observation day.
- Interpretability: Bilirubin, platelets, norepinephrine, and creatinine had the largest mean absolute SHAP values, while nearly all temporal importance was concentrated on Day t−0.Day t−1 contributed a secondary signal for four leading features, and Day t−2 contributed negligibly.
- SOFA component contribution: Cardiovascular dysfunction ranked first for both variance contribution and correlation with total SOFA, at 15.3% variance and r = 0.63.Hepatic, renal, and coagulation dysfunction followed, while neurological and respiratory systems had lower discriminating variation despite higher mean subscores.
- SOFA component contribution: Cardiovascular dysfunction rose from 0.574 in the Low band to 2.755 in the Severe band, while its deterioration-day elevation was largest at +0.546.Neurological and respiratory subscores were already elevated at low severity, leaving less room for escalation.
5 Discussion and Conclusion
The TCN extracted meaningful signal for next-day SOFA prediction, but its short window led it to anchor predictions on recent severity rather than model temporal dynamics. Cardiovascular dysfunction was the strongest discriminator, while complete-case selection constrained performance and generalisability.
- The TCN achieved moderate next-day SOFA accuracy beyond persistence but behaved more like a severity-anchoring model than a true sequence model.SHAP analysis showed predictions were dominated by the most recent observation, limiting the practical value of the three-day window for deterioration detection.
- Cardiovascular, hepatic, coagulation, and renal dysfunction collectively governed cross-sectional severity and longitudinal clinical course in the MIMIC-IV cohort.These systems also differentiated the trajectory groups, linking component analysis with clustering findings.
- Complete-case selection introduced selection bias that constrained model performance at low-severity levels.
Clinically Stratified and Deterioration-Detection Metrics
The regression model tracked stable SOFA trajectories better than abrupt clinical changes, while deterioration detection was strongly biased toward the majority stable class. Representative volatile admissions showed systematic underestimation of rapid deterioration and recovery.
- Deterioration detection: 19 of 733 deteriorating sequences were correctly flagged, while 714 were classified as stable.Among 4,181 stable sequences, 4,147 were correctly identified and 34 were misclassified as deteriorating.
- Deterioration detection: The deteriorating class achieved recall 0.03 and F1-score 0.05, consistent with severe class imbalance favoring stability predictions.The confusion-matrix asymmetry reflects the model’s tendency to predict stability by default.
- Trajectory prediction: Stable SOFA trajectories were tracked relatively closely, whereas sharp deteriorations and recoveries were consistently underestimated or missed.Across six sampled ICU admissions, the trajectories included stable, gradual, and volatile clinical courses.
- Trajectory prediction: Approximately 2–3 points of deterioration were underestimated in the most volatile admission.This smoothing tendency aligns with greater model weighting of the most recent observation than the preceding trajectory.
SOFA Component Contribution
SOFA component contributions varied by severity and by deterioration status. Cardiovascular dysfunction was the strongest severity and acute-worsening discriminator, while respiratory dysfunction fluctuated during deterioration despite low overall variance contribution.
- Subscore distributions: Respiratory and neurological subscores had high medians of 3 with narrow spread, whereas hepatic, coagulation, and renal subscores were zero-inflated.Cardiovascular showed the widest spread, consistent with the binary nature of vasopressor use.
- Severity contribution: Cardiovascular, hepatic, coagulation, and renal contributions rose sharply across Low, Moderate, and Severe SOFA bands.Cardiovascular contribution alone increased nearly fivefold, while neurological and respiratory contributions were already elevated in the Low band.
- Deterioration contribution: Cardiovascular showed the largest deterioration-associated increase, with mean subscore difference of +0.546 between deteriorating and stable days.This was identified as the strongest marker of acute clinical worsening.
- Deterioration contribution: Respiratory ranked second for deterioration-associated increase at +0.398 despite ranking last in variance contribution and correlation.Its acute fluctuation contrasted with its limited contribution to overall severity differences.
Trajectory Clustering
Diagnostic criteria supported a two-cluster trajectory solution rather than a higher number of clusters. The silhouette analysis identified k = 2 as optimal, while the inertia curve lacked a clear higher-k inflection.
- Cluster selection: The silhouette score identified k = 2 as optimal, with a score of 0.456.Scores were assessed across k = 2 to 7.
- Cluster selection: The elbow plot showed smoothly declining within-cluster inertia without a clear inflection point.Neither diagnostic supported selecting more than two clusters.
- Cluster selection: Together, the silhouette and elbow diagnostics supported the two-cluster solution used for trajectory analysis.
SOFA Variable Extraction Reference
Table A.3 documents the SOFA organ-system variables used in the study and maps them to their MIMIC-IV source tables. These variables support the preprocessing and SOFA score calculation described in Section 3.2.
- Table A.3 lists the SOFA organ-system variables used in the study and their corresponding MIMIC-IV source tables.
- The listed variables form the basis of the preprocessing described in Section 3.2.
- The variable-to-source-table mapping supports the SOFA score calculation described in Section 3.2.