📄 medicine

Prognostic Nutritional Index and Machine Learning for Mortality Prediction in Sepsis-Associated Acute Kidney Injury: A MIMIC-IV Cohort Study

This MIMIC-IV cohort study demonstrates that the Prognostic Nutritional Index is an independent, nonlinear predictor of mortality in sepsis-associated acute kidney injury, with its integration into a random forest machine learning model significantly enhancing risk stratification and clinical utility.

Original authors: Pengyan Li, Kai Zhang, Lihua Guo, Xin Liu, Yanpeng Zhang

Published 2026-07-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Pengyan Li, Kai Zhang, Lihua Guo, Xin Liu, Yanpeng Zhang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the human body as a bustling, high-tech city. When a massive storm hits—like a severe infection called sepsis—the city's power grid (the immune system) and its supply lines (nutrition) get knocked out. In this chaos, a specific district often suffers the most: the kidneys, which act as the city's water filtration plant. When the storm damages the filters, it's called sepsis-associated acute kidney injury (S-AKI). Doctors have long used "weather reports" to guess how badly the city will fare, but these reports mostly measure the immediate storm damage, like blood pressure or fever. They often miss the city's long-term resilience: how strong its buildings are and how many repair crews it has left. This is where a simple math formula called the Prognostic Nutritional Index (PNI) comes in. Think of PNI as a "City Resilience Score" that combines two vital stats: the amount of protein in the blood (like the quality of building materials) and the number of immune cells (like the size of the repair crew). The big question researchers wanted to answer was: Does this resilience score actually predict who survives the storm in the kidney district, and can we use a super-smart computer brain to make that prediction even better?

This study, which looked at data from over 4,000 patients in a massive digital hospital database called MIMIC-IV, set out to test if this "City Resilience Score" could help doctors predict who might not make it after 28 days or even a full year. The researchers didn't just look at the score in a straight line; they asked if there was a "sweet spot" where the score was perfect, or if getting too high was also a problem. They also built a team of 15 different computer algorithms—ranging from simple logic to complex "black box" brains—to see which one could best predict the outcome using this score and other data.

The findings were quite revealing. First, the study confirmed that a higher PNI score is generally a good thing. Patients with the highest scores had a 30% lower risk of dying within 28 days compared to those with the lowest scores. It's like having a city with strong buildings and plenty of repair crews; they weather the storm much better. However, the relationship wasn't a simple "higher is always better" line. The researchers discovered a "U-shaped" curve. Imagine a valley: the risk of death is high if your score is too low (the left side of the U), drops to a safe minimum at a score of about 29.13 (the bottom of the valley), and then starts to creep up again if the score gets too high (the right side of the U). This suggests there might be a specific "Goldilocks zone" for this score, rather than just "the more, the merrier."

The study also found that this score is especially important for older patients. For people aged 65 and older, the PNI score was a very strong predictor of survival, acting like a crystal ball for their resilience. For younger patients, the connection was weaker and less clear. To put this all together, the researchers trained a computer model using a "Random Forest" algorithm (think of it as a committee of many decision trees voting on the answer). This model turned out to be the best predictor, with the PNI score ranking as the third most important clue it used, right after the patient's age and a measure of kidney waste called BUN.

The researchers were careful to note that while their computer model worked very well on the data they had, it's a "snapshot" based on information taken only during the first 24 hours of a patient's stay. They suggest that future studies should check if tracking how this score changes over time might tell an even better story. They also pointed out that because this was a look-back study at old data, it shows a strong link but doesn't prove that fixing the score will automatically save lives. However, the results strongly suggest that checking this simple "City Resilience Score" could help doctors spot the most vulnerable patients early, especially the elderly, and perhaps guide them toward better nutrition and care to help their bodies fight off the storm.

Technical Summary: Prognostic Nutritional Index and Machine Learning for Mortality Prediction in Sepsis-Associated Acute Kidney Injury

Problem Statement
Sepsis-associated acute kidney injury (S-AKI) is a critical complication with high morbidity and mortality, often exceeding 40% in-hospital death rates. Current severity scoring systems (e.g., SOFA, SAPS II) primarily capture acute physiological derangements but often fail to reflect the chronic nutritional and immunological reserves that influence patient resilience. While the Prognostic Nutritional Index (PNI)—a composite of serum albumin and total lymphocyte count—has shown prognostic value in other clinical populations, its specific role in S-AKI remains incompletely characterized. Furthermore, existing literature often assumes linear relationships between PNI and mortality, potentially overlooking clinically significant nonlinear patterns or threshold effects. This study addresses these gaps by investigating the association between PNI and mortality in S-AKI patients and developing a machine learning (ML) framework for enhanced risk stratification.

Methodology

Data Source: A retrospective cohort study was conducted using the MIMIC-IV database (2008–2022), comprising over 380,000 ICU admissions.
Study Population: The final analytical cohort included 4,087 adult patients (≥18 years) with a first ICU admission, diagnosed with S-AKI based on Sepsis-3 and KDIGO criteria. Patients with ICU stays <24 hours or missing essential laboratory data (albumin, lymphocyte count) were excluded.
Variable Calculation: PNI was calculated using the formula: $PNI = 10 \times \text{serum albumin (g/dL)} + 0.005 \times \text{total lymphocyte count (/mm}^3)$ .
Statistical Analysis:
- Association: Cox proportional hazards regression (unadjusted and multivariable adjusted) was used to estimate hazard ratios (HR) for 28-day and 365-day all-cause mortality.
- Nonlinearity: Restricted cubic spline (RCS) analysis and two-piecewise Cox models were employed to identify dose-response relationships and threshold effects.
- Subgroup Analysis: Stratified by sex, age, and pulmonary comorbidity, with interaction terms tested via likelihood ratio tests.
Machine Learning Pipeline:
- Feature Selection: The Boruta algorithm and Least Absolute Shrinkage and Selection Operator (LASSO) regression with 10-fold cross-validation were used to identify relevant predictors.
- Model Benchmarking: Fifteen ML algorithms were trained and compared. Performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC), calibration plots, Brier scores, and Decision Curve Analysis (DCA).
- Interpretability: SHapley Additive exPlanations (SHAP) values were used to quantify feature contributions and interpret the "black box" models.

Key Results

Patient Demographics: The cohort had a mean age of 64.9 years, with 57.4% male. Patients in the lowest PNI quartile (Q1) exhibited higher rates of comorbidities (cerebrovascular disease, myocardial infarction), more advanced AKI stages, and higher severity scores compared to the highest quartile (Q4).
Mortality Association: PNI demonstrated a robust inverse association with mortality. In the fully adjusted model, patients in the highest PNI quartile (Q4) had a 30% reduction in 28-day mortality risk (HR 0.70, 95% CI 0.59–0.83) compared to the lowest quartile (Q1). Similar reductions were observed for 365-day mortality.
Nonlinear and Threshold Effects: RCS analysis revealed a significant U-shaped nonlinear relationship with a nadir at approximately PNI 29.13. Two-piecewise Cox models identified breakpoints at 36.42 (28-day) and 35.76 (365-day). Below these thresholds, higher PNI significantly reduced mortality risk; above them, the protective effect diminished or reversed.
Subgroup Findings: The prognostic value of PNI was most pronounced in elderly patients (≥65 years). A significant interaction with age was observed; in patients <65 years, the association was attenuated, whereas in those ≥65 years, the protective effect of high PNI was strong and dose-dependent.
Machine Learning Performance:
- The Random Forest algorithm (via the ranger package) achieved the highest discriminative performance (AUC = 0.773), outperforming CatBoost, Extra Trees, and traditional logistic regression.
- Feature Importance: SHAP analysis ranked PNI as the third most important feature for mortality prediction, following only Age and Blood Urea Nitrogen (BUN).
- Clinical Utility: The model demonstrated excellent calibration and a positive net benefit across a broad range of risk thresholds (0.05–0.90) in Decision Curve Analysis, suggesting clinical utility for risk-stratified management.

Significance and Claims
The paper claims that the Prognostic Nutritional Index serves as an independent, robust prognostic biomarker for all-cause mortality in patients with S-AKI. Its significance lies in three main areas:

Pathophysiological Insight: PNI captures both nutritional status (albumin) and immunocompetence (lymphocytes), reflecting the complex interplay of inflammation and immune paralysis in S-AKI better than traditional scores alone.
Nonlinear Dynamics: The study challenges the assumption of linearity, demonstrating that the relationship between PNI and mortality is U-shaped with specific thresholds, which is critical for defining optimal treatment targets.
Clinical Application: By integrating PNI into a high-performing machine learning model, the study provides a tool for enhanced risk stratification. The authors suggest that routine PNI assessment, particularly in elderly S-AKI patients, could guide therapeutic interventions and resource allocation.

The authors acknowledge limitations, including the retrospective nature of the study, the use of single-timepoint PNI measurements, and the need for external validation in independent cohorts before clinical implementation. They conclude that future prospective studies should validate these thresholds and investigate whether PNI-directed nutritional optimization improves survival.

Technical Summary: Prognostic Nutritional Index and Machine Learning for Mortality Prediction in Sepsis-Associated Acute Kidney Injury

More like this