Machine Learning Prediction of Prolonged Length of Stay in Older Patients with Lung Cancer: A Multicenter Study Using XGBoost with SHAP Interpretation

Abstract

Background

Prolonged length of stay (PLOS) is common among older patients with lung cancer and is associated with increased risks of adverse outcomes, greater healthcare resource utilization, and reduced treatment efficiency. However, reliable risk stratification tools specifically designed for this population, particularly in the Chinese healthcare context, remain lacking.

Methods

This multicenter retrospective study included patients aged 65 years or older with primary lung cancer from two Chinese hospitals. PLOS was defined as a length of stay exceeding the 75th percentile of the overall cohort. After least absolute shrinkage and selection operator‑based feature selection from 34 candidate variables, six machine learning models were developed in the training set and validated using both internal and external validation sets. Model performance was assessed using discrimination, calibration, and clinical utility metrics. The optimal model was interpreted using SHapley Additive exPlanations (SHAP).

Results

A total of 5,836 patients were included. Among the six models, XGBoost achieved the highest discriminative performance, with area under the receiver operating characteristic curve values of 0.821 (95% CI: 0.806–0.837) in the training set, 0.773 in internal validation, and 0.727 in external validation. Calibration curves and decision curve analysis demonstrated satisfactory agreement and net clinical benefit. SHAP analysis identified pneumonia, hypoproteinemia, C‑reactive protein (CRP), uric acid (UA), and admission pathway as the top five predictors. Pneumonia, hypoproteinemia, CRP, admission pathway, and systolic blood pressure were positively associated with PLOS, whereas UA and total protein showed negative associations.

Conclusion

The XGBoost model provided robust and generalizable predictions of PLOS in older Chinese patients with lung cancer. Its interpretability may facilitate early identification of high‑risk individuals and support targeted perioperative interventions in geriatric oncology care.

Introduction

Lung cancer remains the leading cause of cancer-related mortality worldwide, with a median age at diagnosis exceeding 70 years.^1,2 Prolonged length of stay (PLOS) is increasingly recognized as a critical indicator of healthcare efficiency and patient safety in geriatric oncology.³ A retrospective real-world study of 20,650 older cancer patients reported a median length of stay of 9 days, with older age, diabetes, and malnutrition independently associated with PLOS.⁴ Given that older patients frequently present with diminished physiological reserve and multiple comorbidities, they are particularly vulnerable to PLOS.⁵ Timely identification of high-risk individuals using PLOS as a monitoring metric may facilitate early intervention.

PLOS in older adults with lung cancer imposes substantial economic and clinical burdens.⁶ A prospective study demonstrated that older non-small cell lung cancer (NSCLC) patients with cachexia experienced significantly longer cumulative length of stay (80.7 vs 38.5 days) and higher costs (¥3.5 vs ¥2.1 million per person).⁷ A cohort study found that infection-related hospitalization was associated with immediate and persistent cognitive decline among nursing home residents.⁸ In lung cancer, PLOS has been independently associated with increased 30-day readmission rates and worse overall survival.⁹ PLOS also predisposes older patients to a spectrum of hospitalization-associated harms, including nosocomial infections, delirium, falls, and functional decline.¹⁰ With lung cancer accounting for nearly 1.8 million deaths annually worldwide, accurate PLOS prediction is essential for personalized care of this vulnerable population.^11,12

Machine learning (ML) methods have gained traction in predicting PLOS, offering advantages over conventional regression in handling high-dimensional and nonlinear data.¹³ A Japanese study employing five algorithms to predict PLOS in 480 patients undergoing lung cancer surgery found that ridge regression achieved area under the receiver operating characteristic curve (AUC) values of 0.84 and 0.82 in the derivation and test cohorts, respectively.¹⁴ A deep learning study in 953 NSCLC patients achieved an AUC of 0.82 for predicting PLOS.¹⁵ ML has also successfully predicted 5-year survival in older NSCLC (AUC = 0.79) and timeliness of care (AUC = 0.89).^16,17 However, existing models predominantly rely on admission data, lack integration of dynamic in-hospital data, and have limited interpretability (the “black-box” problem). Moreover, prediction tools specifically developed for older Chinese lung cancer patients remain scarce. To address these gaps, this study leverages real-world clinical data from a Chinese cohort to develop and validate multiple ML models for predicting PLOS in older patients with lung cancer. While we acknowledge that our model is also based on static admission data due to the inherent limitations of retrospective data extraction, a key advantage of our study is the deliberate prioritization of clinically accessible baseline features within the first 24 hours, coupled with a strong emphasis on model interpretability via Shapley additive explanations (SHAP). This ensures our model is highly actionable for early risk stratification at the point of admission.

Methods

Study Design and Participants

This was a multicenter retrospective study. We initially collected medical records from patients with lung cancer who received inpatient care at the Affiliated Banan Hospital of Chongqing Medical University between January 2020 and December 2025. After applying predefined eligibility criteria, data from this center were randomly allocated into a training set and an internal validation set at a ratio of 7:3. In addition, we collected data from patients with lung cancer hospitalized at Linfen Central Hospital between January 2018 and December 2025 to serve as an external validation set. The inclusion criterion was a confirmed diagnosis of primary lung cancer requiring inpatient treatment. This broad cohort included both surgical and non-surgical lung cancer patients, encompassing diverse treatment pathways such as surgery, chemotherapy, radiotherapy, and palliative care. Given the primary aim of developing a generalizable bedside risk stratification tool for older patients at admission, we deliberately did not restrict the cohort to specific treatment modalities. The exclusion criteria were as follows: (1) age < 65 years; (2) treatment discontinuation due to in-hospital death or discharge against medical advice; and (3) presence of a severe psychiatric disorder. Ultimately, a total of 5,836 eligible patients were enrolled across the two centers, comprising 3,347 in the training set, 1,435 in the internal validation set, and 1,054 in the external validation set. The detailed participant selection process is illustrated in Figure 1. This study was approved by the ethics committees of both the Affiliated Banan Hospital of Chongqing Medical University (approval no. BNLL-KY-2026-027) and Linfen Central Hospital (approval no. LY2026-35-01), and was conducted in accordance with the Declaration of Helsinki. Owing to the retrospective nature of the study and the use of de-identified data, the requirement for informed consent was waived by the institutional review boards of both participating centers.

Figure 1.

Study participant flow diagram.

Data Collection

A total of 34 potential variables that may influence PLOS in older patients with primary lung cancer were identified, consisting of 27 baseline variables and 7 derived parameters. The 27 baseline variables included age, sex, marital status, admission pathway, smoking history, drinking history, surgical history, diabetes mellitus (DM), hypertension, coronary heart disease (CHD), chronic gastritis, hypoproteinemia, pneumonia, systolic blood pressure (SBP), diastolic blood pressure (DBP), international normalized ratio (INR), D-dimer, mean platelet volume (MPV), platelet distribution width (PDW), white blood cell (WBC) count, C-reactive protein (CRP), blood urea nitrogen (BUN), creatinine (CREA), uric acid (UA), total protein (TP), sodium (Na), and calcium (Ca).

To eliminate potential temporal data leakage, all predictor variables were extracted strictly at a predefined baseline prediction time point: within the first 24 hours of hospital admission. Specifically, diagnoses of pneumonia and hypoproteinemia were based on findings present at admission or documented within the first 24 hours. Laboratory parameters (including CRP, WBC count, D-dimer, etc.) were measured upon admission or within the initial 24-hour window. All variables were available to clinicians at the time of the baseline prediction, and none of the variables reflect events or complications occurring later during the hospitalization. For missing values in quantitative variables, multiple imputation was performed. Specifically, the mice package in R was used to implement predictive mean matching for all quantitative variables with missing data. The specific proportions of missing data for all quantitative variables are detailed in Table S1. A total of m = 5 imputed datasets were generated for each dataset. To ensure the complete isolation of validation data, the imputation model included only the candidate predictor variables, and the imputation process was performed separately and independently within the training set and each validation set.

The remaining seven derived parameters included neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), lymphocyte-to-monocyte ratio (LMR), pan-immune-inflammation value (PIV), prognostic nutritional index (PNI), albumin-to-globulin ratio (AGR), and fibrinogen-to-albumin ratio (FAR). The formulas for PIV and PNI were as follows: PIV = (absolute neutrophil count ×absolute monocyte count × absolute platelet count)/absolute lymphocyte count; PNI = albumin (g/L) + 5 × absolute lymphocyte count (10^9/L). All derived parameters were calculated after multiple imputation was completed.

Definition of PLOS

The primary outcome of this study was the presence of PLOS. Given the lack of a standardized definition for PLOS, it was defined as exceeding the 75th percentile of the training set, in accordance with previous relevant studies.^18–21 Length of stay was defined as the duration from the day of admission to the day of discharge. In this study, PLOS was operationalized as a length of stay greater than 13 days. This specific threshold (>13 days) was exclusively derived from the training set and then applied unchanged to both the internal and external validation sets.

Feature Selection

First, univariate analysis was performed in the training set to identify variables significantly associated with PLOS. Given the relatively high dimensionality of candidate variables (n=34) relative to the number of outcome events in the training cohort, the two-stage strategy was adopted to prioritize clinically stable variables, minimize model overfitting, and ensure parsimonious model construction. Variables showing statistical significance were then entered into the least absolute shrinkage and selection operator (LASSO) regression for further dimension reduction.²² This entire feature selection pipeline was performed once, strictly on the full training set prior to any model training, and was not repeated during the cross-validation process. The final set of variables was locked before proceeding to internal and external validation, ensuring no validation cohort information influenced the variable selection. To address multicollinearity, variance inflation factor (VIF) analysis was conducted on all variables retained by LASSO; no significant multicollinearity was considered to exist when all VIF values were below 5. In addition, a heatmap was generated and correlation coefficients were calculated for continuous variables to illustrate pairwise relationships. Finally, restricted cubic spline (RCS) analysis was employed to explore potential nonlinear associations between continuous variables and PLOS.²³

ML Model Development and Performance Evaluation

Based on the final variables selected by LASSO regression, six different ML models—namely extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), random forest (RF), logistic regression (LR), decision tree (DT), and support vector machine (SVM)—were developed in the training set to predict the risk of PLOS among older patients with lung cancer. Prior to model training, all continuous variables were standardized (Z-score normalization). Importantly, the mean and standard deviation for standardization were calculated exclusively from the training set and then applied to the internal and external validation sets to prevent any information leakage from the validation cohorts. Given the inherent class imbalance (PLOS vs non-PLOS approximately 1:3), the Synthetic Minority Over-sampling Technique (SMOTE) was strictly applied within the training folds during 5-fold cross-validation to ensure that the resampling process did not incorporate validation data. To guarantee reproducibility, a fixed random seed (seed = 42) was set for all procedures. The final optimal values of the hyperparameters for each algorithm are presented in Table S2. Performance metrics for cross‑model comparison included AUC, sensitivity, specificity, Youden index, and F1 score, with AUC serving as the primary measure of discriminative ability. To quantify the precision of these estimates, 95% confidence intervals (CIs) for all performance metrics were calculated using 1,000 bootstrap resampling iterations across all three datasets. To assess calibration fidelity objectively, in addition to visual inspection of calibration curves, the calibration intercept, calibration slope, and Brier score were calculated for all models. To assess calibration fidelity and clinical utility, calibration curves and decision curve analysis (DCA) were performed. To enhance the interpretability of the ML models, the best‑performing model among the six was selected. SHAP methodology was then applied for explanation, including ranking feature importance, calculating SHAP values for each feature, and visualizing these values to clarify the specific associations between features and the risk of PLOS.²⁴

Statistical Analysis

All statistical analyses were performed using R software (version 4.3.3) and Python software (version 3.12.3). Continuous variables following a normal distribution were presented as mean ± standard deviation and compared between groups using Student’s t‑test. Continuous variables with a non‑normal distribution were summarized as median (interquartile range) and compared using the Mann‑Whitney U-test. Categorical variables were expressed as counts (percentages) and analyzed using the chi‑square test. A two‑sided p‑value < 0.05 was considered statistically significant.

Results

Baseline Characteristics of the Study Cohort

A total of 5,836 eligible participants were ultimately included in this study. The median age in both the training and internal validation sets was 71 years (interquartile range: 68–76 years), with males accounting for 72% of the cohort and married individuals representing the vast majority (92.18%) (Table 1). In contrast, the external validation cohort (n = 1,054, from Linfen Central Hospital) exhibited notable clinical differences compared to the derivation cohort (n = 4,782, from the Affiliated Banan Hospital). Specifically, patients in the external cohort were older, presented with a higher proportion of emergency admissions, and had higher baseline levels of CRP and D-dimer. These observed differences reflect the inherent dataset shift across different healthcare settings. Regarding comorbidities, the three most prevalent conditions were hypertension (33.81%), pneumonia (18.03%), and DM (15.85%). The prevalence of PLOS was 25.5% (n = 853) in the training set, 26.1% (n = 375) in the internal validation set, and 34.6% (n = 365) in the external validation set. The median length of stay was 7 days (IQR: 4–13) in both the training set and internal validation set and 10 days (IQR: 7–15) in the external validation set. The results of variable imputation for the training and internal validation sets are presented in Tables S3 and S4, respectively. For most variables, no statistically significant differences were observed before and after imputation (P > 0.05), suggesting a satisfactory imputation performance.

Table 1.

Baseline Characteristics of the Training and Internal Validation Sets

Variables	Total (n = 4,782)	Training Set (n = 3,347)	Internal Validation Set (n = 1,435)	P value
Age (years)	71 (68, 76)	71 (68, 76)	71 (68, 76)	0.631
Sex				0.282
Male	3,443 (72)	2,394 (71.53)	1,049 (73.1)
Female	1,339 (28)	953 (28.47)	386 (26.9)
Marital status				0.868
Married	4,408 (92.18)	3,084 (92.14)	1,324 (92.26)
Unmarried	225 (4.71)	156 (4.66)	69 (4.81)
Divorced/widowed	149 (3.12)	107 (3.2)	42 (2.93)
Admission pathway				0.370
Emergency	2,297 (48.03)	1,593 (47.59)	704 (49.06)
Outpatient	2,485 (51.97)	1,754 (52.41)	731 (50.94)
Smoking history				0.205
No	2,348 (49.1)	1,664 (49.72)	684 (47.67)
Yes	2,434 (50.9)	1,683 (50.28)	751 (52.33)
Drinking history				0.993
No	3,801 (79.49)	2,661 (79.5)	1,140 (79.44)
Yes	981 (20.51)	686 (20.5)	295 (20.56)
Surgical history				0.667
No	3,762 (78.67)	2,627 (78.49)	1,135 (79.09)
Yes	1,020 (21.33)	720 (21.51)	300 (20.91)
DM				0.012
No	4,024 (84.15)	2,846 (85.03)	1,178 (82.09)
Yes	758 (15.85)	501 (14.97)	257 (17.91)
Hypertension				0.360
No	3,165 (66.19)	2,201 (65.76)	964 (67.18)
Yes	1,617 (33.81)	1,146 (34.24)	471 (32.82)
CHD				0.353
No	4,120 (86.16)	2,873 (85.84)	1,247 (86.9)
Yes	662 (13.84)	474 (14.16)	188 (13.1)
Chronic gastritis				0.640
No	4,200 (87.83)	2,945 (87.99)	1,255 (87.46)
Yes	582 (12.17)	402 (12.01)	180 (12.54)
Hypoproteinemia				0.211
No	4,210 (88.04)	2,960 (88.44)	1,250 (87.11)
Yes	572 (11.96)	387 (11.56)	185 (12.89)
Pneumonia				0.332
No	3,920 (81.97)	2,756 (82.34)	1,164 (81.11)
Yes	862 (18.03)	591 (17.66)	271 (18.89)
SBP (mmHg)	126 (115, 138)	126 (115, 138)	126 (115, 138)	0.619
DBP (mmHg)	77 (70, 85)	77 (70, 85)	77 (70, 85)	0.522
INR	1.01 (0.95, 1.07)	1.01 (0.95, 1.07)	1.01 (0.95, 1.07)	0.712
D-dimer (mg/L)	0.51 (0.34, 1.03)	0.51 (0.34, 1.03)	0.52 (0.34, 1.03)	0.926
MPV (fL)	9.4 (8.6, 10.4)	9.4 (8.6, 10.3)	9.5 (8.6, 10.5)	0.125
PDW (%)	16 (15.6, 16.4)	16 (15.6, 16.4)	16 (15.6, 16.3)	0.123
WBC count (×109/L)	6.1 (4.7, 8.2)	6.1 (4.7, 8.1)	6.2 (4.7, 8.4)	0.773
CRP (mg/L)	5.25 (2.93, 19.45)	5.11 (2.95, 18.76)	5.76 (2.89, 21.77)	0.179
BUN (mmol/L)	2.46 (2, 3.08)	2.46 (2, 3.11)	2.46 (1.96, 3.08)	0.711
CREA (μmol/L)	66 (55, 79)	66 (55, 79)	66 (54, 79)	0.481
UA (μmol/L)	323 (262, 388)	322 (262, 386)	326 (263, 389)	0.482
TP (g/L)	67 (63.2, 71)	67 (63.2, 71)	67.2 (63.05, 71)	0.642
Na (mmol/L)	139 (137, 141)	139 (137, 141)	139 (137, 141)	0.944
Ca (mmol/L)	2.24 (2.15, 2.32)	2.24 (2.15, 2.32)	2.23 (2.15, 2.32)	0.382
NLR	4.32 (2.74, 7.08)	4.3 (2.74, 6.96)	4.41 (2.75, 7.37)	0.462
PLR	211.89 (145.96, 315.71)	210.68 (144.44, 314.49)	214.29 (149.3, 318.63)	0.391
LMR	2.26 (1.49, 3.36)	2.26 (1.51, 3.35)	2.24 (1.45, 3.38)	0.686
PIV	396.71 (195.58, 894.38)	395.92 (194.07, 876.48)	397.31 (199.4, 930.14)	0.677
PNI	44.61 (40.9, 47.95)	44.63 (41.05, 48.02)	44.58 (40.61, 47.89)	0.302
AGR	1.42 (1.23, 1.61)	1.43 (1.24, 1.61)	1.41 (1.22, 1.61)	0.097
FAR	0.11 (0.08, 0.14)	0.11 (0.08, 0.14)	0.11 (0.09, 0.14)	0.307

Abbreviations: DM, diabetes mellitus; CHD, coronary heart disease; SBP, systolic blood pressure; DBP, diastolic blood pressure; INR, international normalized ratio; MPV, mean platelet volume; PDW, platelet distribution width; WBC, white blood cell; CRP, C-reactive protein; BUN, blood urea nitrogen; CREA, creatinine; UA, uric acid; TP, total protein; Na, sodium; Ca, calcium; NLR, neutrophil-to-lymphocyte ratio; PLR, platelet-to-lymphocyte ratio; LMR, lymphocyte-to-monocyte ratio; PIV, pan-immune-inflammation value; PNI, prognostic nutritional index; AGR, albumin-to-globulin ratio; FAR, fibrinogen-to-albumin ratio.

LASSO-Driven Variable Selection and Multicollinearity Diagnostics

In the training set, all variables except sex (P = 0.750), marital status (P = 0.055), smoking history (P = 0.358), drinking history (P = 0.342), surgical history (P = 0.334), DBP (P = 0.231), PDW (P = 0.195), and BUN (P = 0.783) showed statistically significant differences between the PLOS and non-PLOS groups (P < 0.05, Table 2). These significantly different variables were then subjected to LASSO regression for feature selection. Under ten-fold cross-validation, the optimal penalty coefficient lambda (lambda.1se = 0.01902102) yielded 13 variables with non-zero coefficients: admission pathway, hypertension, CHD, chronic gastritis, hypoproteinemia, pneumonia, SBP, D-dimer, MPV, WBC count, CRP, UA, and TP (Figure 2).

Table 2.

Comparison of Variables Between PLOS and Non-PLOS Groups in the Training Set

Variables	Total (n = 3,347)	PLOS Group (n = 853)	Non-PLOS Group (n = 2,494)	P value
Age (years)	71 (68, 76)	72 (68, 77)	71 (68, 76)	< 0.001
Sex				0.750
Male	2,394 (71.53)	606 (71.04)	1,788 (71.69)
Female	953 (28.47)	247 (28.96)	706 (28.31)
Marital status				0.055
Married	3,084 (92.14)	770 (90.27)	2,314 (92.78)
Unmarried	156 (4.66)	51 (5.98)	105 (4.21)
Divorced/widowed	107 (3.2)	32 (3.75)	75 (3.01)
Admission pathway				< 0.001
Emergency	1,593 (47.59)	354 (41.5)	1,239 (49.68)
Outpatient	1,754 (52.41)	499 (58.5)	1,255 (50.32)
Smoking history				0.358
No	1,664 (49.72)	412 (48.3)	1,252 (50.2)
Yes	1,683 (50.28)	441 (51.7)	1,242 (49.8)
Drinking history				0.342
No	2,661 (79.5)	668 (78.31)	1,993 (79.91)
Yes	686 (20.5)	185 (21.69)	501 (20.09)
Surgical history				0.334
No	2,627 (78.49)	659 (77.26)	1,968 (78.91)
Yes	720 (21.51)	194 (22.74)	526 (21.09)
DM				0.021
No	2,846 (85.03)	704 (82.53)	2,142 (85.89)
Yes	501 (14.97)	149 (17.47)	352 (14.11)
Hypertension				< 0.001
No	2,201 (65.76)	511 (59.91)	1,690 (67.76)
Yes	1,146 (34.24)	342 (40.09)	804 (32.24)
CHD				< 0.001
No	2,873 (85.84)	657 (77.02)	2,216 (88.85)
Yes	474 (14.16)	196 (22.98)	278 (11.15)
Chronic gastritis				< 0.001
No	2,945 (87.99)	699 (81.95)	2,246 (90.06)
Yes	402 (12.01)	154 (18.05)	248 (9.94)
Hypoproteinemia				< 0.001
No	2,960 (88.44)	621 (72.8)	2,339 (93.79)
Yes	387 (11.56)	232 (27.2)	155 (6.21)
Pneumonia				< 0.001
No	2,756 (82.34)	550 (64.48)	2,206 (88.45)
Yes	591 (17.66)	303 (35.52)	288 (11.55)
SBP (mmHg)	126 (115, 138)	127 (116, 140)	126 (114, 138)	0.003
DBP (mmHg)	77 (70, 85)	77 (71, 85)	77 (70, 85)	0.231
INR	1.01 (0.95, 1.07)	1.03 (0.96, 1.1)	1 (0.95, 1.06)	< 0.001
D-dimer (mg/L)	0.51 (0.34, 1.03)	0.65 (0.39, 1.41)	0.47 (0.33, 0.92)	< 0.001
MPV (fL)	9.4 (8.6, 10.3)	9.7 (8.8, 10.6)	9.4 (8.6, 10.3)	< 0.001
PDW (%)	16 (15.6, 16.4)	16 (15.5, 16.4)	16 (15.7, 16.4)	0.195
WBC count (×109/L)	6.1 (4.7, 8.1)	6.8 (5.4, 9.3)	6 (4.6, 7.8)	< 0.001
CRP (mg/L)	5.11 (2.95, 18.76)	11.35 (3.98, 46.22)	4.45 (2.81, 13.44)	< 0.001
BUN (mmol/L)	2.46 (2, 3.11)	2.46 (1.96, 3.13)	2.46 (2, 3.08)	0.783
CREA (μmol/L)	66 (55, 79)	64 (53, 80)	67 (55, 79)	0.022
UA (μmol/L)	322 (262, 386)	298 (231, 369)	330 (272, 392)	< 0.001
TP (g/L)	67 (63.2, 71)	66 (61, 70)	67.9 (64, 71.3)	< 0.001
Na (mmol/L)	139 (137, 141)	138.1 (135.5, 141)	139.05 (137, 141)	< 0.001
Ca (mmol/L)	2.24 (2.15, 2.32)	2.22 (2.12, 2.31)	2.25 (2.17, 2.33)	< 0.001
NLR	4.3 (2.74, 6.96)	5.15 (3.25, 9.51)	4.04 (2.62, 6.43)	< 0.001
PLR	210.68 (144.44, 314.49)	223.53 (150, 357.3)	207.01 (142.88, 302.85)	< 0.001
LMR	2.26 (1.51, 3.35)	2.03 (1.32, 3.17)	2.33 (1.57, 3.38)	< 0.001
PIV	395.92 (194.07, 876.48)	514.14 (232.01, 1330.77)	372.2 (187.32, 761.09)	< 0.001
PNI	44.63 (41.05, 48.02)	42.76 (38.74, 46.76)	45.16 (41.87, 48.32)	< 0.001
AGR	1.43 (1.24, 1.61)	1.4 (1.18, 1.6)	1.44 (1.25, 1.62)	< 0.001
FAR	0.11 (0.08, 0.14)	0.12 (0.09, 0.16)	0.1 (0.08, 0.13)	< 0.001

Abbreviations: DM, diabetes mellitus; CHD, coronary heart disease; SBP, systolic blood pressure; DBP, diastolic blood pressure; INR, international normalized ratio; MPV, mean platelet volume; PDW, platelet distribution width; WBC, white blood cell; CRP, C-reactive protein; BUN, blood urea nitrogen; CREA, creatinine; UA, uric acid; TP, total protein; Na, sodium; Ca, calcium; NLR, neutrophil-to-lymphocyte ratio; PLR, platelet-to-lymphocyte ratio; LMR, lymphocyte-to-monocyte ratio; PIV, pan-immune-inflammation value; PNI, prognostic nutritional index; AGR, albumin-to-globulin ratio; FAR, fibrinogen-to-albumin ratio.

Figure 2.

Variable selection using LASSO regression. (A) Coefficient shrinkage paths of the thirteen selected predictors. Each curve shows how a variable’s coefficient changes as λ increases. (B) Cross‑validation error as a function of log(λ). Vertical dashed lines indicate the λ value that minimizes the error (left) and the λ value at one standard error from the minimum (right, lambda.1se = 0.01902102).

To thoroughly assess potential multicollinearity among the 13 selected variables, correlation analysis and collinearity diagnostics were performed. As shown in Figure 3, the absolute values of all correlation coefficients between continuous variables did not exceed 0.50. The strongest correlation was observed between WBC count and CRP (r = 0.41), followed by D-dimer and CRP (r = 0.26); the remaining pairs exhibited weak or negligible associations. According to Table S5, VIF values for all variables ranged from 1.025 to 1.241, well below the conventional threshold of 5. Collectively, these findings indicate no substantial multicollinearity among the 13 final predictors.

Figure 3.

Correlation heatmap of continuous variables selected by LASSO. The circles represent the correlation coefficients (r) between variables, with the specific numerical values displayed. The color gradient from red to blue indicates negative to positive correlations, respectively (as shown by the color bar on the right). Notably, the ** symbol inside the circles indicates a statistically significant correlation (P < 0.01).

Dose-Response Relationships of Continuous Variables

To investigate the dose-response relationships between the seven continuous variables selected by LASSO and the risk of PLOS, RCS regression was performed. As shown in Figure 4, SBP (P for nonlinear = 0.076), MPV (P for nonlinear = 0.831), WBC count (P for nonlinear = 0.467), and TP (P for nonlinear = 0.070) exhibited approximately linear dose-response relationships with PLOS risk, with overall associations being statistically significant (all P for overall < 0.001). In contrast, D-dimer (P for nonlinear < 0.001), CRP (P for nonlinear < 0.001), and UA (P for nonlinear < 0.001) demonstrated significant nonlinear associations with PLOS risk. Specifically, the OR for D-dimer increased steeply at lower concentration ranges before reaching a plateau. CRP showed an approximately log-linear positive association, with a more pronounced risk increase at lower concentrations. UA exhibited a U-shaped pattern, with the risk decreasing initially and then increasing as UA levels rose.

Figure 4.

RCS analysis for dose-response relationships. (A) SBP; (B) D-dimer; (C) MPV; (D) WBC count; (E) CRP; (F) UA; (G) TP.

Development and Validation of ML Models

A total of six ML models were constructed and systematically evaluated in the training set, internal validation set, and external validation set. The performance metrics of each model in the training set, including AUC, sensitivity, specificity, Youden index, and F1 score, are summarized in Table 3. Among the six models, XGBoost demonstrated the best overall performance, achieving an AUC of 0.821 (95% CI: 0.806–0.837) in the training set, with corresponding sensitivity, specificity, Youden index, and F1 score of 0.739 (95% CI: 0.709–0.768), 0.781 (95% CI: 0.765–0.796), 0.519 (95% CI: 0.485–0.554), and 0.621 (95% CI: 0.596–0.645), respectively. The DeLong test was employed to compare the AUCs between XGBoost and the other five models across all three datasets. In the training set, XGBoost achieved significantly higher AUCs than LR (P < 0.001), RF (P < 0.001), LightGBM (P < 0.001), and DT (P < 0.001); however, the difference between XGBoost and SVM did not reach statistical significance (P = 0.478). In the internal validation set, XGBoost was significantly superior to DT (P < 0.001) and RF (P < 0.001), but not significantly different from SVM (P = 0.194). Similarly, in the external validation set, XGBoost significantly outperformed DT (P < 0.001), LightGBM (P = 0.017), and SVM (P = 0.013). The performance metrics for the internal and external validation sets are presented in Tables S6 and S7, respectively. The prevalence of PLOS was 25.5% (n = 853) in the training set, 26.1% (n = 375) in the internal validation set, and notably higher at 34.6% (n = 365) in the external validation set, which may partially contribute to the observed decrease in external discriminative performance. XGBoost consistently achieved the highest discrimination among the six models across all datasets, with AUCs of 0.821, 0.773, and 0.727 in the training, internal validation, and external validation sets, respectively. Regarding calibration, the XGBoost model demonstrated excellent agreement between predicted and observed outcomes, with a calibration intercept of −1.057, slope of 1.225, and Brier score of 0.175 in the training set; −1.010, 0.961, and 0.191 in the internal validation set; and −0.916, 1.010, and 0.230 in the external validation set, respectively. The detailed calibration metrics for all models are provided in Table S8.

Table 3.

Performance Metrics of Six ML Models in the Training Set

Model	AUC (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)	Youden Index (95% CI)	F1 Score (95% CI)
XGBoost	0.821 (0.806–0.837)	0.739 (0.709–0.768)	0.781 (0.765–0.796)	0.519 (0.485–0.554)	0.621 (0.596–0.645)
LightGBM	0.805 (0.790–0.822)	0.679 (0.648–0.711)	0.789 (0.774–0.805)	0.468 (0.433–0.504)	0.591 (0.565–0.618)
RF	0.775 (0.758–0.792)	0.655 (0.626–0.689)	0.778 (0.762–0.794)	0.433 (0.398–0.472)	0.569 (0.543–0.596)
LR	0.789 (0.772–0.806)	0.658 (0.626–0.691)	0.782 (0.768–0.799)	0.440 (0.406–0.477)	0.573 (0.547–0.600)
DT	0.771 (0.753–0.790)	0.750 (0.722–0.779)	0.704 (0.687–0.723)	0.454 (0.420–0.489)	0.574 (0.549–0.598)
SVM	0.820 (0.803–0.836)	0.747 (0.717–0.778)	0.736 (0.720–0.753)	0.483 (0.450–0.516)	0.593 (0.569–0.617)

Abbreviations: AUC, area under the ROC curve; CI, confidence interval; XGBoost, extreme Gradient Boosting; LightGBM, light gradient boosting machine; RF, random forest; LR, logistic regression; DT, decision tree; SVM, support vector machine.

All six models were further evaluated through visualization. Figures 5A–C present the ROC curves of the six models for the training, internal validation, and external validation sets, respectively. Figures 6A–C show the corresponding calibration curves, where XGBoost consistently exhibited the closest alignment with the ideal diagonal, indicating better calibration than the other models. Figures 7A–C display the DCA, in which XGBoost demonstrated the highest net benefit across a wide range of threshold probabilities in all three datasets, suggesting superior clinical utility.

Figure 5.

ROC curves of six ML models in the training and validation sets. (A) Training set; (B) Internal validation set; (C) External validation set.

Figure 6.

Calibration curves of six ML models in the training and validation sets. (A) Training set; (B) Internal validation set; (C) External validation set.

Figure 7.

DCA of six ML models in the training and validation sets. (A) Training set; (B) Internal validation set; (C) External validation set.

SHAP-Based Visualization of the Optimal ML Model

In the optimal XGBoost model, we evaluated the key factors influencing the predicted probability of PLOS using both feature importance (Figure 8A) and SHAP summary plots (Figure 8B). Figure 8A identifies the top five predictors: pneumonia, hypoproteinemia, CRP, UA, and admission pathway. To further clarify the direction of their effects, Figure 8B presents the SHAP values, where each dot corresponds to an individual patient, and colors ranging from red to blue represent high to low feature values. The SHAP value on the horizontal axis is positively associated with the model’s predicted probability of PLOS. Notably, features such as pneumonia, hypoproteinemia, CRP, admission pathway, and SBP were positively associated with higher predicted risk of PLOS, whereas UA and TP were negatively associated with predicted risk. It is important to note that these SHAP contributions reflect the importance of these variables within the XGBoost model’s predictive framework and represent predictive associations derived from the model’s internal decision-making process. They do not imply independent or causal associations with PLOS.

Figure 8.

SHAP-based interpretation of the XGBoost model. (A) Feature importance bar chart; (B) Summary dot plot of SHAP values. Red and blue colors indicate high and low feature values, respectively. SHAP values to the left of the vertical axis reduce the predicted risk of PLOS, whereas values to the right increase it.

Discussion

In this multicenter retrospective study of 5,836 older Chinese patients with primary lung cancer, we developed and externally validated six ML models to predict PLOS. Among these, the XGBoost model demonstrated the highest discriminative performance across all three datasets, achieving AUCs of 0.821, 0.773, and 0.727 in the training, internal, and external validation sets, respectively. The moderate external discrimination (AUC = 0.727) and the observed performance decay from the training set reflect the inherent challenges in generalizing prediction models across different healthcare settings and patient populations. Therefore, while the model offers clinically acceptable risk stratification, its transportability requires further prospective refinement and validation in broader cohorts. Our findings align with several recent investigations demonstrating the utility of ML methods for PLOS prediction in lung cancer populations. A study of 6,767 lung cancer patients undergoing video-assisted thoracoscopic surgery reported that RF achieved an AUC of 0.792 for predicting postoperative PLOS, with surgical duration and age emerging as the most critical features.²⁵ Similarly, an explainable ML framework using the MIMIC-III dataset reported that RF achieved an AUC of 0.98 for lung cancer length of stay prediction after addressing class imbalance.²⁶ More recently, a study of 105 lung cancer patients undergoing pulmonary resection identified total lung capacity, maximal inspiratory pressure, and peak expiratory flow as independent predictors of PLOS, together accounting for 47% of the variance in PLOS.²⁷ Notably, the present study provides external validation—a critical but often omitted step in predictive modeling research—thereby strengthening the evidence for model generalizability beyond the derivation institution.

Beyond model development and comparison, our study leveraged RCS regression to explore dose-response relationships between seven continuous predictors and PLOS risk, revealing nuanced non-linear patterns that would have been obscured by conventional linear modeling. RCS analysis demonstrated that D-dimer, CRP, and UA exhibited significant non-linear associations with PLOS (all P for nonlinearity < 0.001). Specifically, CRP showed an approximately log-linear positive association, with a more pronounced risk increase at lower concentration ranges—a pattern consistent with findings from a study of gastrointestinal cancer patients, where CRP levels at ICU admission demonstrated a non-linear relationship with prolonged ICU stay, with an inflection point at 2.6 (log-transformed mg/L).²⁸ Additionally, a study in internal medicine wards found that each 80 mg/L increment in admission CRP was associated with a 2.09-fold increased risk of PLOS.²⁹ The U-shaped relationship observed for UA in our study—where PLOS risk decreased initially and then increased with rising UA levels—suggests a more complex biological interplay that warrants further investigation. This biphasic pattern is consistent with emerging evidence from cardiovascular research. A large cohort study of 33,096 patients examining long-term stable serum UA levels reported a U-shaped association between UA and all-cause mortality in both sexes, with both low and high UA levels conferring elevated risk.³⁰ Converging evidence from a study of 1,109 geriatric patients with respiratory failure further showed that elevated serum UA levels correlated with PLOS and increased ventilation requirements.³¹ These findings underscore the methodological advantage of RCS over traditional approaches that assume linearity, as falsely imposing linear constraints may lead to underestimation of risk at extreme biomarker values or failure to identify protective ranges.

Among the 13 predictors selected by LASSO regression, pneumonia emerged as the most important feature in the XGBoost model, followed by hypoproteinemia and CRP. The predominance of pneumonia as a predictor is clinically plausible, as respiratory infections in older patients with lung cancer often trigger a cascade of adverse events including respiratory failure, sepsis, and functional decline, each contributing independently to PLOS.³² This finding is further supported by recent thoracic surgical evidence demonstrating that a higher comorbidity burden significantly predicts postoperative pulmonary complications in patients undergoing lung cancer surgery, thereby prolonging recovery and hospitalization.³³ However, it must be emphasized that these SHAP-based findings represent predictive associations within our model and should be interpreted cautiously, rather than establishing independent risk factors or causal relationships. The clinical utility of these markers lies in their ability to flag high-risk patients for early intervention, but they should not be interpreted as definitive etiological pathways without further confirmatory studies. Hypoproteinemia, reflecting poor nutritional status, has been consistently associated with delayed recovery and increased complication risk in geriatric oncology populations.³⁴ A narrative review on nutritional support in lung cancer surgery concluded that hypoalbuminemia and reduced body mass index are strongly associated with increased postoperative complications and PLOS, emphasizing the prognostic importance of preoperative nutritional assessment.³⁵ The elevated CRP levels observed in high-risk patients in our cohort corroborate evidence from patients with resected NSCLC, where postoperative CRP levels at week 6 were independently associated with worse overall survival and recurrence-free survival (HR 2.23, 95% CI not reported, P < 0.001), reinforcing CRP’s role as a prognostic biomarker reflecting systemic inflammation in lung cancer.³⁶ Collectively, these findings suggest that the triad of active infection (pneumonia), nutritional depletion (hypoproteinemia), and systemic inflammation (CRP) forms a synergistic risk constellation for PLOS in older lung cancer patients.

Beyond identifying risk factors, the ultimate value of a prediction model lies in its ability to inform actionable clinical decisions. Our XGBoost-based model, coupled with SHAP visualization, offers individualized risk estimates that may potentially guide future targeted interventions for older lung cancer patients at high risk of PLOS. For instance, patients flagged with high predicted probability—particularly those exhibiting the triad of pneumonia, hypoproteinemia, and elevated CRP—might be prioritized for enhanced nutritional support, early infectious workup, or closer respiratory monitoring in future clinical practice. Conversely, in patients with favorable feature profiles (eg, low CRP, normal UA, elective admission), a streamlined care pathway with reduced observation intensity might be safely considered, thereby optimizing resource allocation in bed-constrained settings. However, it is important to emphasize that these proposed interventions (including nutritional support, infection management, respiratory monitoring, and discharge planning) have not been tested in this retrospective study, and should be considered as potential future applications. This risk-stratified approach aligns with the principles of precision medicine in geriatric oncology, where one-size-fits-all protocols are increasingly replaced by individualized care plans.³⁷ Nevertheless, future prospective clinical-impact studies are necessary to determine whether model-guided care actually reduces PLOS incidence or improves patient outcomes.

Nevertheless, several barriers must be addressed before clinical deployment. First, seamless integration with electronic health record systems is technically challenging but essential for real-time risk calculation without disrupting clinical workflows.³⁸ Second, clinician acceptance and training are critical, as SHAP outputs—while interpretable—require basic data literacy to be correctly acted upon. Third, prospective implementation studies are needed to determine whether model-assisted risk stratification actually reduces PLOS incidence or merely identifies high-risk patients without improving outcomes. A recent cluster randomized trial of a ML-based early warning system for sepsis demonstrated that providing risk scores alone did not improve outcomes without accompanying decision support and protocolized responses.³⁹ Thus, our model should be embedded within a broader intervention bundle—including pre-admission optimization for elective cases, early geriatric consultation for high-risk admissions, and structured discharge planning—rather than deployed as a standalone prediction tool.

Several limitations warrant consideration. First, the retrospective design inherently carries the risk of selection bias and unmeasured confounding, despite our efforts to include a comprehensive set of clinical variables. Second, all data were derived from two Chinese hospitals; although external validation was performed, broader generalization to other healthcare systems or ethnic populations requires further prospective, multicenter validation. Third, PLOS was defined based on the 75th percentile of the cohort (13 days), which may limit comparability with studies adopting different cutoffs, and the lack of a universally accepted definition remains a challenge in this field. Fourth, dynamic in-hospital parameters, such as daily changes in inflammatory markers or treatment-related events, were not incorporated into the current models, potentially omitting time-varying influences that could further refine real-time risk prediction. Fifth, we acknowledge that crucial oncological determinants of length of stay, including cancer stage, histological subtype, treatment modality (eg, surgery, chemotherapy, or radiotherapy), performance status, and specific admission indications, were not available in our retrospectively extracted database. Although our cohort included a mixed population of surgical and non-surgical patients, we were unable to adjust for these treatment-related factors. The absence of these clinically significant variables may limit the interpretability of our model regarding why certain patients are at risk, and its transportability to centers with different treatment protocols. Future prospective studies should integrate these essential oncological parameters to enhance the model’s clinical utility and generalizability.

Conclusions

In conclusion, older patients with lung cancer face a substantially elevated risk of PLOS, which in turn predisposes them to a cascade of adverse outcomes and strains healthcare resources. The XGBoost model developed in this study offers a promising and externally validated tool with moderate discrimination for early identification of high-risk individuals in the Chinese geriatric population. By translating electronic health record data into an actionable risk estimate with transparent feature contributions, this approach may hold potential for future integration into clinical decision support systems to guide targeted perioperative and supportive care strategies. However, the clinical utility and impact of this model on patient outcomes remain to be established through prospective implementation and impact assessment studies. Future efforts should focus on validating the model in broader, more diverse cohorts and evaluating whether model-assisted risk stratification genuinely improves patient-centered and economic outcomes in real-world geriatric oncology settings.

Funding Statement

This study was funded by Linfen Central Hospital (LFSZXYY-KJK-YJ-2026006).

Data Sharing Statement

The datasets used and/or analyzed in the current study are available from the corresponding authors upon reasonable request.

Ethics Approval and Consent to Participate

This study was approved by the ethics committees of both the Affiliated Banan Hospital of Chongqing Medical University (approval no. BNLL-KY-2026-027) and Linfen Central Hospital (approval no. LY2026-35-01). Because the study employed a retrospective design, informed consent in written form was not deemed necessary, and this requirement was formally waived by the ethics committee. The entire study was conducted in strict observance of the ethical standards set forth in the Declaration of Helsinki (1964) as well as its subsequent amendments. Furthermore, all methodologies and procedures strictly conformed to pertinent ethical guidelines and statutory regulations.

Author Contributions

All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

All authors declare no conflictS of interest.