A Prediction Model for Persistent Inflammation-Immunosuppression-Catabolism Syndrome in Patients with Sepsis: An Ambispective Cohort Study Authors: Xiangning Zhong, Jindai Yang, Le Chang, Xiaoyu Wang, Lan Peng, Wenjuan Yan, Hongbin Hu, Zhongqing Chen, Yaoyuan Zhang Published: 2026 ABSTRACT ObjectiveTo develop and temporally validate an admission-variable model for estimating the risk of persistent inflammation-immunosuppression-catabolism syndrome (PICS) among patients with sepsis who remained in the intensive care unit (ICU) for ≥14 days.MethodsWe conducted an ambispective, single-center cohort study. The retrospective development cohort included patients admitted from January 2023 to May 2025, and the prospective temporal validation cohort included patients admitted from June 2025 to February 2026. Candidate predictors measured at ICU admission were selected using least absolute shrinkage and selection operator (LASSO) regression and entered into a multivariable logistic regression model. Discrimination, calibration, and potential clinical utility were evaluated using the area under the receiver operating characteristic curve (AUC), calibration analyses, and decision curve analysis.ResultsThe development cohort included 242 patients, of whom 93 (38.4%) met the PICS criteria. Compared with respiratory infection, urinary infection (adjusted odds ratio (aOR 0.07, 95% CI 0.01–0.67), gastrointestinal infection (aOR 0.15, 95% CI 0.05–0.45), and skin or soft-tissue infection (aOR 0.29, 95% CI 0.11–0.77) were associated with lower odds of PICS. Higher lactate was associated with higher odds of PICS (aOR 1.20, 95% CI 1.07–1.34), whereas higher lymphocyte count (aOR 0.46, 95% CI 0.28–0.77), albumin (aOR 0.92, 95% CI 0.86–0.99), and vitamin D (aOR 0.88, 95% CI 0.83–0.94) were associated with lower odds. The AUC was 0.803 (95% CI 0.747–0.859) in the development cohort and 0.791 (95% CI 0.718–0.864) in the temporal validation cohort.ConclusionAmong patients with sepsis who remained in the ICU for ≥14 days, an admission-variable model showed acceptable discrimination in the development and temporal validation cohorts. External multicenter validation and prospective impact evaluation are required before clinical use. Introduction Sepsis is life-threatening organ dysfunction caused by a dysregulated host response to infection and remains a major cause of morbidity and mortality among hospitalized patients.1,2 Improvements in guideline-based care have increased short-term survival.3 Shifting attention toward chronic critical illness (CCI), persistent organ dysfunction, and prolonged ICU dependence among survivors.4,5 Persistent inflammation–immunosuppression–catabolism syndrome (PICS) has been proposed as a phenotype underlying CCI and is characterized by sustained inflammation, immune dysfunction, catabolism, muscle wasting, and recurrent infection.6 Observational studies suggest that PICS may occur in 30–50% of critically ill patients7 and is associated with substantial long-term mortality.8 Research has progressed from conceptual definitions toward mechanistic profiling, including emergency myelopoiesis, myeloid-derived suppressor cell expansion, persistent immune suppression, and catabolism,9,10 However, PICS is generally identified only after a prolonged ICU course, using clinical and laboratory criteria assessed around ICU day 14.6 Existing prognostic models use different populations and predictor sets, and their transportability remains uncertain.11–13 A pragmatic model based on routinely available admission variables may therefore complement, rather than replace, existing approaches. Vitamin D deficiency is common during critical illness and is associated with altered immune function, systemic inflammation, and skeletal-muscle dysfunction.14–21 Lower 25-hydroxyvitamin D concentrations have also been associated with greater illness severity and worse outcomes,22,23 although randomized trials of early high-dose supplementation have been neutral overall.24 Its value as a noncausal risk marker for subsequent PICS remains insufficiently studied. We therefore aimed to develop and temporally validate a pragmatic admission-variable model for estimating PICS risk among patients with sepsis who remained in the ICU for ≥14 days. Materials and Methods Study Design and Participants This ambispective cohort study was conducted in the ICU of Nanfang Hospital (Guangzhou, China). The retrospective development cohort included patients admitted between January 2023 and May 2025. The prospective temporal validation cohort included patients admitted between June 2025 and February 2026 and used the same eligibility criteria. We screened adults with sepsis defined according to Sepsis-3.0 criteria.1 Because PICS was assessed on ICU day 14, the study population was restricted to patients who remained in the ICU for ≥14 days. We excluded patients with >20% missingness in critical clinical variables, advanced malignancy, previous organ transplantation, chemotherapy or radiotherapy within 30 days before admission, chronic immunosuppressant or glucocorticoid use, end-stage renal disease (CKD stage 5), end-stage hepatic disease (Child–Pugh class C), pregnancy, or lactation. The study protocol, including retrospective medical-record review and prospective data collection, was approved by the Ethics Committee of Nanfang Hospital, Southern Medical University (Ethical Approval No. NFEC-2025-127) on March 28, 2025. The requirement for informed consent was waived by the Ethics Committee for both the retrospective and prospective cohorts. All data were de-identified before analysis, and the study was conducted in accordance with the Declaration of Helsinki. Definitions of PICS PICS was defined using modified criteria derived from the original description by Gentile et al6 Patients were classified as having PICS if they remained in the ICU for ≥14 days and met all of the following laboratory criteria on ICU day 14: (1) C-reactive protein >1.5 mg/L; (2) peripheral blood lymphocyte count <0.8 × 109/L; and (3) either (a) serum albumin <30 g/L, regardless of albumin supplementation, or (b) serum albumin 30–35 g/L despite albumin supplementation before the day-14 measurement. Patients with albumin 30–35 g/L who had not received albumin supplementation were not considered to meet the albumin criterion. Because this modified albumin criterion depends partly on treatment exposure, its reproducibility should be evaluated in external cohorts using prespecified supplementation definitions. Data Collection Data were extracted from the electronic medical record. Variables recorded at ICU admission included demographic characteristics, comorbidities, primary infection source, laboratory measurements, Acute Physiology and Chronic Health Evaluation II score, Sequential Organ Failure Assessment score, early enteral nutrition, and use of continuous renal replacement therapy. Outcomes included duration of mechanical ventilation within 28 days, hospital and ICU length of stay, 28-day mortality, and in-hospital mortality. Laboratory measurements used to determine PICS status—C-reactive protein, lymphocyte count, and serum albumin—were obtained on ICU day 14. Statistical Analysis Continuous variables are presented as mean ± standard deviation for approximately normally distributed data or median (interquartile range) otherwise. Groups were compared using the independent-samples t test or Mann–Whitney U-test, as appropriate. Categorical variables are presented as number (percentage) and were compared using the χ2-test or Fisher exact test. Missing data were handled using multiple imputation by chained equations (mice package, version 3.18.0). The imputation model included all candidate predictors and the outcome. Five datasets were generated using predictive mean matching, and estimates were pooled according to Rubin’s rules. A complete-case sensitivity analysis produced qualitatively consistent results. Candidate predictors were selected using least absolute shrinkage and selection operator (LASSO) regression. Variables retained by LASSO were entered into a multivariable logistic regression model to estimate adjusted odds ratios (aORs) with 95% CIs and construct a nomogram. Nested cross-validation with 10 outer folds and 10 inner folds was used to evaluate the LASSO-tuning, variable-selection, and logistic-refitting pipeline. In each outer iteration, the penalty parameter was selected by inner cross-validation using the outer training data, and performance was evaluated in the held-out outer fold. The final model was then fitted using the full development cohort. For temporal validation, the fixed coefficients from the development model were applied to the validation cohort without refitting. Discrimination was evaluated using receiver operating characteristic curves and AUCs with 95% CIs. Calibration was examined using bootstrap calibration plots and the Hosmer–Lemeshow test, and potential clinical utility was explored using decision curve analysis. SHAP (Shapley additive explanations) values were calculated for predictions from the final multivariable logistic regression model and visualized using the shapviz package (version 0.10.3). Global mean absolute SHAP values were interpreted as model-based contribution rankings and were not assumed to reproduce coefficient or nomogram rankings. All tests were two-sided, with P<0.05 considered statistically significant. Analyses were performed using R version 4.5.1. Results During the development period, 275 ICU patients with sepsis were screened; 242 met the eligibility criteria and were included in the final analysis (Figure 1). During the temporal validation period (June 2025–February 2026), 203 patients were screened and 149 met the same eligibility criteria after 54 exclusions (reasons shown in Figure S1). Baseline characteristics of the development and temporal validation cohorts are presented in Table S1. Figure 1Flow chart of the study.A flowchart of sepsis patient cohort selection and PICS risk model development. Flow chart of the study. In the development cohort, compared with the non-PICS group, the PICS group had a higher median SOFA score (8.0 vs 6.0, P=0.006) and more frequently received continuous renal replacement therapy (53.76% vs 38.93%, P=0.024). The PICS group also had higher median lactate (2.20 vs 1.50 mmol/L, P=0.001), lower vitamin D (9.80 vs 13.96 ng/mL, P<0.001), lower lymphocyte count (0.54 vs 0.83 × 109/L, P<0.001), and lower serum albumin (28.40 vs 29.40 g/L, P=0.004). Respiratory infection was more frequent in the PICS group, although the difference was not statistically significant (53.8% vs 40.9%, P=0.054), and the median duration of mechanical ventilation was longer (12.0 vs 9.0 days, P=0.023) (Table 1).Table 1Demographic and Clinical Characteristics of the Development CohortVariablesNon-PICS Group (n = 149)PICS Group (n = 93)PAge, years60.0 ± 15.361.2 ± 14.80.533BMI, kg/m222.86 [20.76, 25.21]22.49 [20.76, 25.06]0.158Gender [n (%)]0.969Male107 (71.8)67 (72.0)Female42 (28.2)26 (28.0)SOFA score6.00 [5.00, 9.00]8.00 [6.00, 11.00]0.006APACHE II score26.00 [22.00, 28.00]26.00 [23.00, 30.00]0.177MV [n (%)]109 (73.2)77 (82.8)0.084Ventilator days M (IQR)9.0 (0.0, 17.0)12.0 (5.0 21.0)0.023CRRT [n (%)]58 (38.9)50 (53.8)0.024EEN [n (%)]35 (23.5)15 (16.1)0.169Comorbidities [n (%)]Diabetes55 (36.9)27 (29.0)0.208Cancer12 (8.1)15 (16.1)0.052CKD17 (11.4)15 (16.1)0.292Hepatic insufficiency11 (7.4)7 (7.5)0.967Hypertension62 (41.6)33 (35.5)0.342Source Of Infection [n (%)]0.054 Respiratory system61 (40.9)50 (53.8) Urinary9 (6.0)2 (2.2) Gastrointestinal29 (19.5)10 (10.8) Skin/soft tissue27 (18.1)11 (11.8) Other23 (15.4)20 (21.5)Laboratory parameters M (IQR)Lac (mmol/L)1.50 [1.10, 2.30]2.20[1.30, 3.70]0.001VD (ng/mL)13.96 [10.31, 17.86]9.80 [7.34, 12.94]<0.001White blood cell (×109/L)10.89 [7.25, 16.39]10.92 [7.57, 16.07]0.797LYM (×109/L)0.83 [0.53, 1.22]0.54 [0.34, 0.71]<0.001LYM at ICU day 14 (×109/L)1.15 [0.88, 1.62]0.47 [0.36, 0.64]<0.001Na+ (mmol/L)141.00 [136.00, 144.00]141.00 [137.00, 147.00]0.393BUN (mmol/L)11.00 [7.60, 19.60]11.50 [7.40, 20.90]0.451Cr (umol/L)119.00 [71.00, 237.00]128.00 [75.00, 273.00]0.554ALB (g/L)29.40 [27.00, 33.30]28.40 [26.00, 30.30]0.004ALB at ICU day 14 (g/L)34.00 [30.10, 36.30]29.80 [28.10, 32.20]<0.001CRP (mg/L)106.87 [59.88, 173.20]121.18 [58.18, 190.15]0.658CRP at ICU day 14 (mg/L)45.55 [21.34, 74.88]60.56 [37.70, 104.61]0.004BNP (pg/mL)2494.00 [544.00, 6294.00]2807.00 [941.00, 7455.00]0.420PCT (ng/mL)2.60 [0.53, 25.00]4.43 [1.26, 13.70]0.295IL-6 (pg/mL)80.00 [31.90, 255.00]103.00 [52.40, 380.00]0.043OutcomeICU Length of Stay (days)19.0 [14.0, 27.0]21.00 [15.0, 29.0]0.185Hospital Length of Stay (days)28.0 [21.0, 38.0]28.0 [21.0, 39.0]0.78728-day mortality, n (%)9 (6.04)11 (11.83)0.112In-hospital mortality, n (%)7 (4.70)8 (8.60)0.220Abbreviations: BMI, Body Mass Index; MV, Mechanical Ventilation; M (IQR), Median (Interquartile Range); SOFA, Sequential Organ Failure Assessment; CRRT, Continuous Renal Replacement Therapy; APACHE II, Acute Physiology and Chronic Health Evaluation II; EEN, Early enteral Nutrition; CKD, Chronic Kidney Disease; Lac, Lactate; VD, Vitamin D; LYM, Lymphocyte Count; Na+, Sodium; BUN, Blood Urea Nitrogen; Cr, Creatinine; ALB, Albumin; CRP, C-reactiveprotein; BNP, B-type Natriuretic Peptide; PCT, Procalcitonin; IL-6, Interleukin-6; ICU, Intensive Care Unit. Table 1Demographic and Clinical Characteristics of the Development CohortVariablesNon-PICS Group (n = 149)PICS Group (n = 93)PAge, years60.0 ± 15.361.2 ± 14.80.533BMI, kg/m222.86 [20.76, 25.21]22.49 [20.76, 25.06]0.158Gender [n (%)]0.969Male107 (71.8)67 (72.0)Female42 (28.2)26 (28.0)SOFA score6.00 [5.00, 9.00]8.00 [6.00, 11.00]0.006APACHE II score26.00 [22.00, 28.00]26.00 [23.00, 30.00]0.177MV [n (%)]109 (73.2)77 (82.8)0.084Ventilator days M (IQR)9.0 (0.0, 17.0)12.0 (5.0 21.0)0.023CRRT [n (%)]58 (38.9)50 (53.8)0.024EEN [n (%)]35 (23.5)15 (16.1)0.169Comorbidities [n (%)]Diabetes55 (36.9)27 (29.0)0.208Cancer12 (8.1)15 (16.1)0.052CKD17 (11.4)15 (16.1)0.292Hepatic insufficiency11 (7.4)7 (7.5)0.967Hypertension62 (41.6)33 (35.5)0.342Source Of Infection [n (%)]0.054 Respiratory system61 (40.9)50 (53.8) Urinary9 (6.0)2 (2.2) Gastrointestinal29 (19.5)10 (10.8) Skin/soft tissue27 (18.1)11 (11.8) Other23 (15.4)20 (21.5)Laboratory parameters M (IQR)Lac (mmol/L)1.50 [1.10, 2.30]2.20[1.30, 3.70]0.001VD (ng/mL)13.96 [10.31, 17.86]9.80 [7.34, 12.94]<0.001White blood cell (×109/L)10.89 [7.25, 16.39]10.92 [7.57, 16.07]0.797LYM (×109/L)0.83 [0.53, 1.22]0.54 [0.34, 0.71]<0.001LYM at ICU day 14 (×109/L)1.15 [0.88, 1.62]0.47 [0.36, 0.64]<0.001Na+ (mmol/L)141.00 [136.00, 144.00]141.00 [137.00, 147.00]0.393BUN (mmol/L)11.00 [7.60, 19.60]11.50 [7.40, 20.90]0.451Cr (umol/L)119.00 [71.00, 237.00]128.00 [75.00, 273.00]0.554ALB (g/L)29.40 [27.00, 33.30]28.40 [26.00, 30.30]0.004ALB at ICU day 14 (g/L)34.00 [30.10, 36.30]29.80 [28.10, 32.20]<0.001CRP (mg/L)106.87 [59.88, 173.20]121.18 [58.18, 190.15]0.658CRP at ICU day 14 (mg/L)45.55 [21.34, 74.88]60.56 [37.70, 104.61]0.004BNP (pg/mL)2494.00 [544.00, 6294.00]2807.00 [941.00, 7455.00]0.420PCT (ng/mL)2.60 [0.53, 25.00]4.43 [1.26, 13.70]0.295IL-6 (pg/mL)80.00 [31.90, 255.00]103.00 [52.40, 380.00]0.043OutcomeICU Length of Stay (days)19.0 [14.0, 27.0]21.00 [15.0, 29.0]0.185Hospital Length of Stay (days)28.0 [21.0, 38.0]28.0 [21.0, 39.0]0.78728-day mortality, n (%)9 (6.04)11 (11.83)0.112In-hospital mortality, n (%)7 (4.70)8 (8.60)0.220Abbreviations: BMI, Body Mass Index; MV, Mechanical Ventilation; M (IQR), Median (Interquartile Range); SOFA, Sequential Organ Failure Assessment; CRRT, Continuous Renal Replacement Therapy; APACHE II, Acute Physiology and Chronic Health Evaluation II; EEN, Early enteral Nutrition; CKD, Chronic Kidney Disease; Lac, Lactate; VD, Vitamin D; LYM, Lymphocyte Count; Na+, Sodium; BUN, Blood Urea Nitrogen; Cr, Creatinine; ALB, Albumin; CRP, C-reactiveprotein; BNP, B-type Natriuretic Peptide; PCT, Procalcitonin; IL-6, Interleukin-6; ICU, Intensive Care Unit. Demographic and Clinical Characteristics of the Development Cohort Abbreviations: BMI, Body Mass Index; MV, Mechanical Ventilation; M (IQR), Median (Interquartile Range); SOFA, Sequential Organ Failure Assessment; CRRT, Continuous Renal Replacement Therapy; APACHE II, Acute Physiology and Chronic Health Evaluation II; EEN, Early enteral Nutrition; CKD, Chronic Kidney Disease; Lac, Lactate; VD, Vitamin D; LYM, Lymphocyte Count; Na+, Sodium; BUN, Blood Urea Nitrogen; Cr, Creatinine; ALB, Albumin; CRP, C-reactiveprotein; BNP, B-type Natriuretic Peptide; PCT, Procalcitonin; IL-6, Interleukin-6; ICU, Intensive Care Unit. LASSO regression selected eight predictors from 26 candidate variables: body mass index, SOFA score, lactate, vitamin D, lymphocyte count, albumin, cancer, and source of infection (Figure S2). No strong pairwise correlations were observed among these predictors (all absolute correlation coefficients <0.3; Figure S3). In the multivariable model (Table 2), higher lactate was associated with higher odds of PICS, whereas higher lymphocyte count, albumin, and vitamin D were associated with lower odds. Compared with respiratory infection, urinary, gastrointestinal, and skin or soft-tissue infections were associated with lower odds of PICS. The model is presented as a nomogram in Figure 2.Table 2Univariate and Multivariable Logistic Regression Analyses of Variables Selected by LASSO RegressionVariablesUnivariate Logistic RegressionMultivariable Logistic RegressionOR (95% CI)PaOR (95% CI)PCancer2.20 (0.98 ~ 4.93)0.0571.67 (0.65 ~ 4.30)0.291Source Of Infection Respiratory1.00 (Reference)1.00 (Reference) Urinary0.27 (0.06 ~ 1.31)0.1050.07 (0.01 ~ 0.67)0.020 Gastrointestinal0.42 (0.19 ~ 0.95)0.0360.15 (0.05 ~ 0.45)<0.001 Skin/soft tissue0.50 (0.22 ~ 1.10)0.0850.29 (0.11 ~ 0.77)0.013 Other1.06 (0.52 ~ 2.15)0.8700.68 (0.29 ~ 1.56)0.360BMI0.93 (0.85 ~ 1.01)0.0720.97 (0.87 ~ 1.07)0.500Lac1.11 (1.01 ~ 1.22)0.0301.20 (1.07 ~ 1.34)0.002SOFA1.10 (1.02 ~ 1.18)0.0091.06 (0.97 ~ 1.15)0.172LYM0.45 (0.26 ~ 0.77)0.0030.46 (0.28 ~ 0.77)0.003ALB0.92 (0.87 ~ 0.97)0.0050.92 (0.86 ~ 0.99)0.036VD0.89 (0.84 ~ 0.94)<0.0010.88 (0.83 ~ 0.94)<0.001Abbreviations: OR, Odds Ratio; aOR, Adjusted Odds Ratio; CI, Confidence Interval; BMI, Body Mass Index; Lac, Lactate; SOFA, Sequential Organ Failure Assessment; LYM, Lymphopenia; ALB, Albumin; VD, Vitamin D. Figure 2The nomogram to predict the risk of PICS in sepsis patients who remained in the ICU for ≥14 days. For the variable “source of infection”, the numerical coding corresponds to the following categories: 1 = Respiratory, 2 = Urinary, 3 = Gastrointestinal, 4 = Skin/soft tissue, 5 = Other.A nomogram infographic for estimating PICS risk in sepsis by assigning points to multiple clinical variables and converting the sum to risk. Reading order is top to bottom: a Points scale from 0 to 100, then variable rows, then outcome scales. Variable rows are labeled, cancer with values 0 and 1; source of infection with values 1, 2, 3, 4, 5; BMI (kg/m superscript 2) from 22 to 36; SOFA from 2 to 16; Lac (mmol/L) from 0 to 35; VD (ng/mL) from 0 to 35; LYM (times 10 superscript 9/L) from 0 to 10; ALB (g/L) from 10 to 50. Some rows show increasing numbers left to right, while VD, LYM and ALB show decreasing numbers left to right. Below the variables, Total Points ranges from 0 to 260. Total Points maps to Linear Predictor from minus 8 to 4 and then to Risk with labeled values 0.1, 0.3, 0.5, 0.7, 0.9.A nomogram scoring chart assigning points to clinical variables to estimate PICS risk in sepsis. Table 2Univariate and Multivariable Logistic Regression Analyses of Variables Selected by LASSO RegressionVariablesUnivariate Logistic RegressionMultivariable Logistic RegressionOR (95% CI)PaOR (95% CI)PCancer2.20 (0.98 ~ 4.93)0.0571.67 (0.65 ~ 4.30)0.291Source Of Infection Respiratory1.00 (Reference)1.00 (Reference) Urinary0.27 (0.06 ~ 1.31)0.1050.07 (0.01 ~ 0.67)0.020 Gastrointestinal0.42 (0.19 ~ 0.95)0.0360.15 (0.05 ~ 0.45)<0.001 Skin/soft tissue0.50 (0.22 ~ 1.10)0.0850.29 (0.11 ~ 0.77)0.013 Other1.06 (0.52 ~ 2.15)0.8700.68 (0.29 ~ 1.56)0.360BMI0.93 (0.85 ~ 1.01)0.0720.97 (0.87 ~ 1.07)0.500Lac1.11 (1.01 ~ 1.22)0.0301.20 (1.07 ~ 1.34)0.002SOFA1.10 (1.02 ~ 1.18)0.0091.06 (0.97 ~ 1.15)0.172LYM0.45 (0.26 ~ 0.77)0.0030.46 (0.28 ~ 0.77)0.003ALB0.92 (0.87 ~ 0.97)0.0050.92 (0.86 ~ 0.99)0.036VD0.89 (0.84 ~ 0.94)<0.0010.88 (0.83 ~ 0.94)<0.001Abbreviations: OR, Odds Ratio; aOR, Adjusted Odds Ratio; CI, Confidence Interval; BMI, Body Mass Index; Lac, Lactate; SOFA, Sequential Organ Failure Assessment; LYM, Lymphopenia; ALB, Albumin; VD, Vitamin D. Univariate and Multivariable Logistic Regression Analyses of Variables Selected by LASSO Regression Abbreviations: OR, Odds Ratio; aOR, Adjusted Odds Ratio; CI, Confidence Interval; BMI, Body Mass Index; Lac, Lactate; SOFA, Sequential Organ Failure Assessment; LYM, Lymphopenia; ALB, Albumin; VD, Vitamin D. The nomogram to predict the risk of PICS in sepsis patients who remained in the ICU for ≥14 days. For the variable “source of infection”, the numerical coding corresponds to the following categories: 1 = Respiratory, 2 = Urinary, 3 = Gastrointestinal, 4 = Skin/soft tissue, 5 = Other. The model had an apparent AUC of 0.803 (95% CI 0.747–0.859) in the development cohort (Figure 3A) and an AUC of 0.791 (95% CI 0.718–0.864) in the temporal validation cohort (Figure 3B). The nested cross-validation procedure yielded a mean out-of-sample AUC of 0.724 (SD 0.115). The full model showed greater apparent discrimination than any individual predictor and than models restricted to traditional PICS diagnostic variables (Figure S4). Based on global mean absolute SHAP values, infection source and vitamin D had the largest model-based contribution rankings (Figure 4). These rankings reflect both fitted effects and the distributions of the predictors and should not be interpreted as causal importance. Figure 3Receiver operating characteristic curves of development cohort (A) and validation cohort (B). The area under the curve (AUC) for each variable in panel (A) is as follows: BMI 0.554, cancer 0.540, source of infection 0.602, SOFA 0.604, ALB 0.609, Lac 0.624, VD 0.691, LYM 0.702, and the Full Model 0.803. In the validation set (B), the AUC values are: Lac 0.525, BMI 0.545, cancer 0.548, ALB 0.586, source of infection 0.608, SOFA 0.621, VD 0.681, LYM 0.688, and the Validation Model 0.791.Image A and B depict multi-line ROC curve plots with axes labeled Specificity and Sensitivity. Both plots feature a diagonal reference line from (1.0, 0.0) to (0.0, 1.0). Image A includes curves for BMI, cancer, infection source, SOFA, ALB, Lac, VD, LYM and Full Model, with AUC values: BMI 0.554, cancer 0.540, infection source 0.602, SOFA 0.604, ALB 0.609, Lac 0.624, VD 0.691, LYM 0.702, Full Model 0.803. The Full Model curve is highest above the diagonal, while BMI and cancer are closer to it. Image B features curves for Lac, BMI, cancer, ALB, infection source, SOFA, VD, LYM and Validation Model, with AUC values: Lac 0.525, BMI 0.545, cancer 0.548, ALB 0.586, infection source 0.608, SOFA 0.621, VD 0.681, LYM 0.688, Validation Model 0.791. The Validation Model curve is highest above the diagonal, while Lac is closer to it. Curves are distinguished by colors and are step-shaped.A multi-line ROC curve graph comparing predictors and full or validation models across two cohorts. Figure 4Ranking of Variables Importance Using SHAP Values.Horizontal bar chart comparing eight predictors, with infection source and VD having the longest bars. Receiver operating characteristic curves of development cohort (A) and validation cohort (B). The area under the curve (AUC) for each variable in panel (A) is as follows: BMI 0.554, cancer 0.540, source of infection 0.602, SOFA 0.604, ALB 0.609, Lac 0.624, VD 0.691, LYM 0.702, and the Full Model 0.803. In the validation set (B), the AUC values are: Lac 0.525, BMI 0.545, cancer 0.548, ALB 0.586, source of infection 0.608, SOFA 0.621, VD 0.681, LYM 0.688, and the Validation Model 0.791. Ranking of Variables Importance Using SHAP Values. Calibration plots for the development and temporal validation cohorts are shown in Figure S5. The Hosmer–Lemeshow tests did not detect statistically significant lack of fit (development cohort: χ2=6.293, P=0.614; validation cohort: χ2=10.782, P=0.214), although these tests alone do not establish good calibration. Decision curves for both cohorts are presented in Figure S6; their clinical usefulness should be interpreted within the threshold-probability ranges displayed in those plots. Discussion In this ambispective, single-center cohort of adults with sepsis who remained in the ICU for ≥14 days, we developed and temporally validated a pragmatic model for estimating PICS risk. The model included eight routinely available admission variables and had an apparent AUC of 0.803 in the development cohort and an AUC of 0.791 in the temporal validation cohort. These findings support further external evaluation of admission-based risk stratification in this selected population but do not establish clinical effectiveness. The biological basis of PICS is multifactorial and incompletely defined. Contemporary models emphasize emergency myelopoiesis and persistent expansion of myeloid-derived suppressor cells, together with ongoing tissue injury, dysregulated inflammation, adaptive immune suppression, and sustained catabolism.6,25 Recurrent infection and continued release of danger- and pathogen-associated molecular patterns may reinforce immune–metabolic dysfunction.26,27 Although we did not measure myeloid-derived suppressor cells directly, patients with PICS had marked lymphopenia and higher IL-6 at admission (Table 1), findings consistent with post-sepsis immune dysregulation. A recent review of sepsis-induced vasoplegia described vascular smooth-muscle phenotypic switching and vascular remodeling as possible contributors to persistent vascular dysfunction.28 These mechanisms may be relevant to prolonged organ dysfunction, but our study did not assess vascular remodeling; therefore, this literature provides biological context rather than direct mechanistic evidence for PICS in our cohort. In the SHAP ranking, source of infection was the most influential contributor. Respiratory infection was associated with the highest risk of PICS compared with other infection sites, even after adjustment for acute illness severity (SOFA score and lactate). Prior studies have reported worse outcomes among patients with sepsis of respiratory origin.29–32 Disruption of the alveolar–capillary barrier may facilitate systemic translocation of pathogens and inflammatory mediators, while achieving early source control in the lung is often challenging.33–36 In addition, prolonged mechanical ventilation increases the risk of ventilator-associated lung injury and secondary infections, potentially sustaining inflammatory signaling relevant to PICS biology.35,37 Consistent with this framework, the PICS group had a higher proportion of respiratory infections and longer ventilation duration than the non-PICS group. Admission vitamin D concentration was inversely associated with subsequent PICS. Vitamin D signaling has immunological and skeletal-muscle effects that are biologically relevant to PICS,14–21 and deficiency is common during critical illness. However, randomized trials of early high-dose vitamin D in critically ill patients have been neutral overall,24,38 and sepsis-focused meta-analyses remain uncertain.39 Moreover, acute illness can lower circulating 25-hydroxyvitamin D through redistribution, hemodilution, and reduced vitamin D-binding protein concentrations.40,41 Admission vitamin D should therefore be interpreted as a severity-associated risk marker in this study, not as evidence of a causal or readily modifiable mechanism. We did not measure vitamin D-binding protein, free 25-hydroxyvitamin D, or detailed supplementation exposure, and residual confounding is likely. Lymphocyte count was another independent predictor and represents a core clinical manifestation of sepsis-associated immunosuppression. Persistent antigen exposure, T-cell exhaustion and apoptosis, impaired antigen presentation, and bone marrow dysfunction contribute to sustained lymphopenia and reduced effector function during and after sepsis.6,8,18,42 In line with prior reports linking lymphopenia to sepsis severity,43 patients who developed PICS had substantially lower lymphocyte counts at admission, supporting lymphopenia as an early marker of vulnerability to persistent immune dysregulation. Hypoalbuminemia is widely associated with poor outcomes in critical illness and reflects the combined effects of systemic inflammation, capillary leak, impaired hepatic synthesis, and nutritional deficits.44–47 Albumin also participates in redox balance and immune modulation, and low concentrations often coincide with malnutrition-related immune impairment.48–51 In our cohort, albumin levels at admission and on ICU day 14 were lower in patients who developed PICS and remained independently associated with PICS risk. However, its relatively smaller contribution in SHAP ranking suggests that albumin may primarily capture downstream inflammatory and catabolic burden rather than act as a specific driver. Elevated lactate, classically interpreted as a marker of hypoperfusion and stress metabolism, was independently associated with PICS in our cohort. Emerging data suggest that lactate can also shape immune responses through immunometabolic signaling, including macrophage polarization and modulation of inflammatory pathways.52–54 Lactate has been implicated in promoting MDSC expansion,55 a putative mechanism in PICS pathobiology.6,25 While causality cannot be established in this study, our findings are consistent with the hypothesis that early metabolic stress may accompany—and potentially reinforce—subsequent immune dysfunction. Although the SOFA score captures acute organ dysfunction, it did not retain independent significance in the multivariable model and had limited incremental importance in SHAP ranking. This suggests that acute physiologic derangement at admission may be insufficient to explain later development of persistent immune–metabolic dysregulation once metabolic stress and host-reserve markers are considered. Lactate and SOFA are correlated and may provide overlapping prognostic information.56–58 In addition, exploratory studies have linked vitamin D receptor–related signaling with severity indices, including SOFA and lactate.59 The coexistence of these variables in the final model may therefore reflect intersecting pathways connecting acute severity, immunometabolism, and longer-term trajectories. BMI and cancer history were retained in the final model despite not reaching independent statistical significance, consistent with the feature-selection properties of LASSO. Clinically, both variables plausibly reflect baseline host reserve: low BMI may signal pre-existing nutritional depletion, and malignancy can be associated with chronic immune alterations and treatment-related immunosuppression.60,61 Larger studies are needed to clarify their independent contributions and potential interactions with immune–metabolic markers. We did not observe statistically significant differences in short-term outcomes (ICU or hospital length of stay, 28-day mortality, or in-hospital mortality) between groups. However, patients who developed PICS exhibited persistent lymphopenia, hypoalbuminemia, and higher CRP by ICU day 14, consistent with the construct that PICS reflects a chronic trajectory of persistent inflammation, immunosuppression, and catabolism rather than early mortality.6 The clinical burden of this phenotype is therefore more likely to emerge as prolonged functional dependence, recurrent infections, and late adverse outcomes. Future studies should incorporate post-discharge endpoints (eg, readmission, long-term mortality, and functional status) to quantify the impact of early risk stratification and targeted supportive interventions. Several limitations warrant consideration. First, this was a single-center study, and external validation in independent multicenter cohorts is required before clinical implementation. Second, the final model contained 11 parameter degrees of freedom (six continuous predictors, one binary predictor, and four indicator variables for the five-level infection-source variable). With 93 PICS events, the events-per-parameter value was 93/11=8.45, indicating a meaningful risk of overfitting and coefficient instability. LASSO selection followed by unpenalized logistic refitting does not eliminate this risk, and the difference between the apparent AUC (0.803) and nested cross-validated AUC (0.724) should be considered when interpreting performance. Third, residual confounding is likely. We did not measure monocyte HLA-DR, cytokine panels, myeloid-derived suppressor cells, vitamin D-binding protein, free 25-hydroxyvitamin D, or detailed vitamin D supplementation, and we could not quantify milder baseline renal or hepatic impairment. Fourth, because PICS was determined on ICU day 14, the analysis was conditional on remaining in the ICU for ≥14 days and does not apply to patients who died, were discharged, or were transferred earlier. Fifth, the modified albumin component of the PICS definition partly depends on albumin supplementation and may reduce reproducibility across centers with different treatment practices. Finally, predictors were measured at a single admission time point. Larger external studies should evaluate model transportability, calibration, the robustness of the outcome definition, and prospective clinical impact. Conclusion This study developed a pragmatic admission-variable model for PICS risk estimation among patients with sepsis who remained in the ICU for ≥14 days. Development and temporal validation showed acceptable discrimination, but external multicenter validation and prospective impact assessment are required before clinical implementation.