Adding discharge characteristics to improve six-month post-discharge mortality prediction in under-five children with suspected sepsis in Ugandan hospitals

This study demonstrates that incorporating just three discharge characteristics into risk prediction models significantly improves the accuracy of six-month post-discharge mortality predictions for under-five children with suspected sepsis in Ugandan hospitals compared to models relying solely on admission data.

The Gist

Imagine a hospital as a busy airport terminal and sick children as travelers who have just finished a rough flight (their illness). The doctors' job is to decide when these travelers are ready to leave the terminal and go home.

For a long time, the airport staff had a rulebook for deciding who could leave. They looked at how the traveler looked when they first arrived at the airport (the Admission data). If the traveler looked okay back then, they were given a "green light" to go home.

The Problem:
The problem is that a traveler might look okay when they arrive, but their luggage might break, or they might get a stomach ache during their stay in the terminal. The old rulebook didn't check for these new problems before letting them leave. As a result, some children who looked "okay" on paper were actually still very fragile, and sadly, many of them got sick again or passed away once they got home.

The New Solution: The "Exit Check" Upgrade
This paper is about upgrading the airport's rulebook. The researchers, working in Ugandan hospitals, decided to add a final "Exit Check" before sending children home.

Instead of just looking at how the child was when they arrived, they also looked at how the child was doing right before they left. They added just three simple questions to the final checklist:

1. Can they breathe well? (Oxygen levels)
2. Can they eat/drink well? (Feeding status)
3. How are they leaving? (Are they going home normally, being sent to a bigger hospital, or leaving against medical advice?)

The Analogy of the "Weather Forecast"
Think of the old model as a weather forecast made on Monday morning. It might predict sunny skies for the weekend. But if a storm rolls in on Friday, that Monday forecast is useless for the weekend.

The new model is like a live weather app that updates right before you step outside. It takes the Monday forecast and adds the current storm clouds. This gives a much more accurate picture of whether it's safe to go out.

What They Found
The researchers tested this new "Exit Check" on over 8,800 children. Here is what happened:

• Better Accuracy: The new model was much better at spotting the children who were actually in danger. It's like upgrading from a blurry security camera to a high-definition one.
• Fewer False Alarms: The old model sometimes sounded the alarm for kids who were actually fine (false positives). This meant nurses had to waste time and resources checking on kids who didn't need it. The new model stopped sounding the alarm for those safe kids.
• Saving Resources: By correctly identifying the "safe" kids, the hospital can focus its limited energy, money, and staff on the kids who are truly at risk. It's like a fire department focusing on the burning houses instead of checking every house on the block.

The Result
By adding just these three simple checks at the very end of the hospital stay, the doctors could predict who might die within six months of going home much more accurately.

• For babies under 6 months, the prediction accuracy went up by 5%.
• For older kids (6 months to 5 years), it went up by 4%.

Why This Matters
In places with limited resources, you can't check on every single child who leaves the hospital. You need a smart way to pick the ones who really need help. This new "Exit Check" acts like a smart filter. It helps doctors say, "This child is strong and ready to go home," and "This child is still fragile; we need to make a special plan to keep them safe."

In short, this paper shows that checking the child's condition right before they leave is just as important as checking them when they arrive. It's a simple tweak that could save many lives by ensuring the right kids get the right care at the right time.

Technical Summary

1. Problem Statement

In Low- and Middle-Income Countries (LMICs), a significant proportion of children under five die after being discharged from the hospital, particularly following treatment for suspected sepsis. Current risk prediction models used in settings like Uganda (e.g., the "Smart Discharge" program) rely exclusively on admission characteristics.

• Limitation: Admission-only models fail to account for the child's clinical trajectory, response to therapy, in-hospital complications, or physiological status at the time of discharge.
• Consequence: This leads to suboptimal risk stratification, potentially resulting in unnecessary follow-up for low-risk survivors or missed interventions for high-risk children who appear stable at admission but deteriorate or fail to recover fully during hospitalization.
• Goal: To develop and validate updated prediction models that integrate discharge-specific characteristics to improve the accuracy of six-month post-discharge mortality predictions.

2. Methodology

Study Design and Data Source

• Type: Secondary analysis of prospective cohort data.
• Setting: Six hospitals in Uganda (Mbarara, Holy Innocents, Masaka, Jinja, Villa Maria, and Uganda Martyrs) between 2012 and 2021.
• Population: 8,810 children under five admitted with suspected sepsis.
  • Cohort 1: 3,665 children aged <6 months.
  • Cohort 2: 5,145 children aged 6–60 months.
• Outcome: Mortality within 183 days (6 months) of discharge.
• Follow-up: 97.8% (under-6m) and 98.2% (6-60m) of participants had outcome data available.

Model Development Strategy
The study utilized an Elastic Net regression approach, which combines Lasso and Ridge regression to handle correlated predictors and perform variable selection.

1. Baseline Models (Admission-Only): Existing models (Model A for <6m, Model D for 6-60m) using admission variables.
2. Intermediate Models: Developed by adding 5 prespecified discharge variables to the admission variables:
   • Oxygen saturation ($SpO_2$)
   • Feeding status
   • Discharge status (routine, referred, or unplanned/DAMA)
   • Respiratory rate
   • Length of hospital stay (LOS)
3. Final Reduced Models: To ensure parsimony and feasibility in low-resource settings, the final models were restricted to the original admission variables plus the top three discharge variables based on Variable Importance (VI) ranking.
   • Model C (<6m): Admission variables + $SpO_2$ at discharge, Discharge status, Feeding status.
   • Model F (6-60m): Admission variables + $SpO_2$ at discharge, Discharge status, Feeding status.

Statistical Analysis

• Validation: 10-fold cross-validation.
• Metrics: Area Under the Receiver Operating Characteristic Curve (AUROC), Brier score (calibration), Precision-Recall AUC (PR-AUC), and Net Reclassification Index (NRI).
• Thresholding: A probability threshold was set to achieve 80% sensitivity for comparison.
• Handling Missing Data: Multiple imputation was used during development.

3. Key Contributions

• Integration of Discharge Data: Demonstrated that adding just three discharge variables significantly enhances predictive power over admission-only models.
• Variable Selection: Identified that $SpO_2$ at discharge, Discharge status (specifically unplanned discharge or referral), and Feeding status are the most critical discharge predictors across both age groups.
• Reclassification Improvement: Showed that the primary benefit of the new models is the reduction of False Positives (correctly identifying low-risk survivors as low-risk), which is crucial for optimizing limited healthcare resources.
• Practical Implementation: The models are designed to be lightweight, requiring only a few additional data points available at the point of discharge, making them suitable for integration into Electronic Health Records (EHR) or mobile health tools in resource-limited settings.

4. Results

Performance Metrics (Comparison: Admission-Only vs. Discharge-Augmented)

Metric	<6 Months Cohort	6–60 Months Cohort
Model	Model A (Admission) vs. Model C (Augmented)	Model D (Admission) vs. Model F (Augmented)
AUROC	Improved from 0.77 to 0.82 ($P<0.001$)	Improved from 0.75 to 0.81 ($P=0.0004$)
Brier Score	0.06 (Consistent improvement in accuracy)	0.04 (Consistent improvement in accuracy)
PR-AUC	Improved from 0.24 to 0.37	Improved from 0.17 to 0.28
NRI	10.41% (Net improvement)	14.51% (Net improvement)

Key Findings on Reclassification:

• The Net Reclassification Index (NRI) was driven almost entirely by the correct reclassification of survivors from "High Risk" to "Low Risk."
  • <6m: 14.97% of survivors were correctly downgraded to low risk.
  • 6-60m: 19.53% of survivors were correctly downgraded to low risk.
• This reduction in false positives means fewer children would be subjected to unnecessary, resource-intensive follow-up interventions.
• Calibration: The augmented models showed superior calibration, particularly at higher risk levels, compared to admission-only models.

Variable Importance:

• For both age groups, the top discharge predictors were consistently $SpO_2$ at discharge, Discharge Status (unplanned/referred), and Feeding Status.
• In the <6m group, "Referral to higher level of care" and "Unplanned discharge" were highly significant.
• In the 6-60m group, "Unplanned discharge" and "Not feeding at all" were critical.

5. Significance and Implications

• Clinical Utility: The study validates that re-evaluating risk at the moment of discharge captures the "evolving risk profile" of the child, which admission data misses. This allows for more precise targeting of post-discharge interventions (e.g., community health worker visits, phone follow-ups).
• Resource Optimization: By reducing false positives, the healthcare system can avoid wasting limited resources on low-risk children who are unlikely to die post-discharge, focusing instead on the truly vulnerable.
• Policy Impact: The findings support the integration of discharge-specific algorithms into national health strategies (like Uganda's Smart Discharge program) to reduce the high burden of post-discharge mortality.
• Future Directions: The authors note the need for external validation in other LMIC settings and the development of lightweight imputation strategies for point-of-care use where data may be missing. They also highlight the growing importance of pulse oximetry access in these settings.

Conclusion:
Adding a minimal set of discharge characteristics ($SpO_2$, feeding, and discharge status) to existing admission-based models significantly improves the discrimination, calibration, and reclassification accuracy of six-month post-discharge mortality predictions for Ugandan children with suspected sepsis. This approach offers a scalable, data-driven method to optimize post-discharge care and reduce child mortality.