This secondary analysis examined older dysphagia patients receiving percutaneous endoscopic gastrostomy (PEG) or total parenteral nutrition (TPN) at a Japanese center (2014–2017). Study data were sourced from Dryad Digital Repository (doi: 10.5061/dryad.gg407h1) Miyanomori Memorial Hospital's Ethics Board waived informed consent and approved resource utilization, as the study utilized publicly available and anonymized data, obviating the need for individual patient consent.

Nutritional support modality (PEG/TPN) was determined through clinician-patient/family consensus based on clinical evaluation. Exclusion criteria comprised: (1) advanced cancer; (2) PEG for gastric decompression; (3) pre-2014 PEG placement.

The study encompassed four main components: (1) Demographic variables: age and gender; (2) Nutritional variables: clinical frailty scale (CFS), body mass index (BMI), and nutritional support (NS) type, nutrient intake; (3) Comorbidities: cardiovascular diseases (CVD), neuromuscular diseases (NMD), ischemic heart diseases (IHD), chronic heart failure (CHF), chronic pulmonary diseases (CPD), chronic liver diseases (CLD), and chronic kidney diseases (CKD); (4) laboratory biomarkers measured within 7 days prior to nutritional support initiation: albumin (ALB), total lymphocyte count (TLC), C-reactive protein (CRP), total cholesterol (TC), and hemoglobin (Hb).

The CALLY index was determined using the following formula, adapted from prior research: CALLY index = [albumin level (g/dl) × lymphocyte count (cells/μl)]/[CRP level (mg/dl) × 10⁴]. Given its right-skewed distribution (Supplementary Figure 1), CALLY values underwent log transformation (Ln-CALLY) for regression analysis. Patients were stratified into two groups based on Ln-CALLY values using maximally selected rank statistics (Figure 1A): a low group (Ln-CALLY ≤ 2.21) and a high group (Ln-CALLY > 2.21).

The primary outcome was all-cause mortality, defined as the time from initiation of nutritional support to death, irrespective of etiology. The secondary outcomes were pneumonia-cause mortality, defined as the duration from the onset of nutritional support to pneumonia-associated death, oral recovery, discharge to home, the incidence of severe pneumonia and sepsis.

Patient characteristics are summarized overall and by Ln-CALLY (low vs. high). Continuous variables are reported as medians [interquartile ranges (IQRs)], and categorical variables as frequencies (%). Continuous data were analyzed using the Wilcoxon rank-sum test. Categorical data were analyzed using Pearson's chi-square test or Fisher's exact test, as appropriate. To ensure data completeness and increase the accuracy of results, multiple imputation methods were used to handle missing values using a chained equation approach method and 5 replications to address the missing data (“mice” package).

Restricted cubic splines (knots at 10th, 50th, and 90th percentiles; reference, Ln-CALLY = 2.21) explored the dose–response relationship between Ln-CALLY and all-cause mortality, with non-linearity tested by the likelihood-ratio test (P for non-linear). The log-rank test and Kaplan–Meier survival curves were used to compare all-cause and pneumonia-related mortality between the two Ln-CALLY groups. To assess Ln-CALLY's independent predictive value, we established multivariate Cox proportional hazards regression models with three models for confounder adjustment. Model 1 was adjusted for age and gender. Model 2 was adjusted for age, gender, NS, BMI, CFS and nutrient intake. Model 3 incorporated all Model 2 variables plus severe dementia, CLD, CKD and TC. Schoenfeld residuals confirmed the proportional-hazards assumption for every covariate (all P > 0.05); any future violations will be addressed with time-varying covariates or stratified analyses.

To assess the robustness of our findings, we conducted sensitivity analyses using the median and tertiles of the Ln-CALLY cut-off, and various imputation methods including predictive mean matching (PMM), random-forest imputation, complete-case analysis, single-stochastic normal imputation, and classification and regression threes (CART). Additionally, we performed direct comparisons of CALLY's predictive performance (e.g., ROC/AUC) against the prognostic nutritional index (PNI).

Stratification and interaction analyses were also performed to evaluate the effects of variables such as age (<85 and ≥85 years), gender, NS (PEG and TPN), severe dementia, CVD, NMD, CHF, CLD and CKD. Forest plots were used to present the results. A two-tailed test with P < 0.05 was considered statistically significant. All statistical analyses were conducted using R (version 4.4.3).

Our analysis encompassed 253 elderly patients with dysphagia, of whom 154 (61.0%) were female. The patients had a median age 85.0 years (IQR: 79.0–89.0) and BMI 19.1 kg/m² (IQR: 17.1–20.8). Their median nutrient intake was 900.0 Kcal/day (IQR: 892.0–930.0). Of the 253 patients, 180 underwent PEG feeding and 73 received TPN. The all-cause and pneumonia-related mortality rates were 55% and 17%, respectively. Oral recovery and sepsis occurred in 5.9% and 15% of patients, severe pneumonia in 5.9%, and 15% were discharged home.

The median Ln-CALLY values were 0.4 (IQR: −0.2–2.8) for the low group and 3.3 (IQR: 2.7–4.2) for the high group. Median Ln-CALLY values were 0.4 (IQR −0.2–2.8) in the low group vs. 3.3 (IQR 2.7–4.2) in the high group. Baseline characteristics (Table 1) were analyzed across the Ln-CALLY groups. High Ln-CALLY group patients demonstrated a greater likelihood of PEG utilization, increased daily nutrient intake, and a higher prevalence of CVD. Conversely, this group exhibited a lower prevalence of severe dementia and CHF. Significant disparities were noted between the groups in terms of ALB, TLC count, CRP, TC, CCI, and Hb levels (all P < 0.05). Relative to the low group, high-group patients demonstrated significantly reduced rates of severe pneumonia, sepsis, all-cause mortality, and pneumonia-related mortality, alongside higher rates of oral intake recovery and home discharge.

Univariate analysis demonstrated reduced severe pneumonia/sepsis rates and improved oral recovery/home discharge rates in high vs. low groups.

An RCS regression model with four knots and a cutoff of Ln-CALLY = 2.21 (reference) revealed a linear association between Ln-CALLY and all-cause mortality (P for overall <0.001, P for non-linear = 0.302) (Figure 1B). Kaplan-Meier analysis showed significant differences in all-cause (P < 0.001) and pneumonia-cause mortality (P = 0.0025) across Ln-CALLY groups (Figure 2).

Cox proportional hazards regression models (Table 2) indicated Ln-CALLY was a significant all-cause mortality predictor in Model 1 (HR: 0.77; 95% CI: 0.70–0.85; P < 0.001), Model 2 (HR: 0.79; 95% CI: 0.72–0.88; P < 0.001), and Model 3 (HR: 0.84; 95% CI: 0.76–0.94; P = 0.002) as a continuous variable, with comparable results for pneumonia-cause mortality (all P < 0.01).

As a categorical variable, high Ln-CALLY showed significantly lower all-cause mortality vs. low group. This finding persisted across Cox models: Model 1 (HR: 0.34; 95% CI: 0.22–0.53; P < 0.001), Model 2 (HR: 0.43; 95% CI: 0.27–0.68; P < 0.001), and Model 3 (HR: 0.46; 95% CI: 0.29–0.72; P < 0.001). Multivariate Cox regression analyses for pneumonia-cause mortality showed a similar trend (P < 0.05).

To further validate our findings, we conducted sensitivity analyses. We recalculated the primary endpoint using the median and upper/lower tertiles in addition to the data-driven threshold identified by maximally selected rank statistics. The hazard ratios remained nearly identical across all cut-points (Supplementary Figure 3), indicating that the observed effects were robust and not dependent on the specific threshold chosen.

We also refitted the multivariable Cox model under five different scenarios for handling missing data: predictive mean matching, random-forest imputation, complete-case analysis, single-stochastic normal imputation, and CART. The point estimates and 95% confidence intervals showed substantial overlap (Supplementary Figure 4), confirming that our conclusions were consistent and insensitive to the underlying missing data mechanism.

To assess the predictive accuracy of Ln-CALLY and PNI for all-cause mortality at the 2-year timepoint, we conducted ROC curve analyses. The area under the curve (AUC) was 0.689 (95% CI: 0.597–0.782) for Ln-CALLY and 0.684 (95% CI: 0.598–0.771) for PNI, demonstrating very similar performance between the two markers.

To further explore the relationship between Ln-CALLY and all-cause mortality, stratified analyses were conducted based on age (<85 years or ≥85 years), gender, NS, presence of severe dementia, CVD, NMD, CHF, CLD, and CKD (Figure 3). Compared to patients in the low group, only those in the high group who underwent TPN or had severe dementia, NMD, CLD, or CKD showed no significant difference in all-cause mortality (all P > 0.05). Notably, significant interactions were observed between Ln-CALLY levels and three clinical parameters: NS (P for interaction = 0.021), CVD (P for interaction = 0.018), and NMD (P for interaction = 0.033).

In the subgroup analysis of pneumonia-cause mortality, patients in the high group showed a significantly lower mortality risk than those in the low group in specific subgroups: those aged <85 years, males, those not receiving TPN, and those without severe dementia, CLD, or CKD, as well as those with CVD and CHF (all P < 0.05). However, no significant interaction effects were found across these subgroups (P for interaction > 0.05) (Figure 4).