Development and multicenter validation of a dynamic nomogram for early postoperative neurocognitive disorder in older adults with hip fracture

Abstract

Background

Perioperative neurocognitive disorder (PND) is a serious complication in older adults following hip fracture surgery, associated with poor functional recovery and increased mortality. Existing prediction models often focus solely on in-hospital delirium, neglecting early post-discharge cognitive decline. This study aimed to develop and externally validate a dynamic nomogram to predict early PND, defined as cognitive impairment occurring within 3 months postoperatively.

Methods

This multicenter, retrospective cohort study included patients aged ≥ 65 years undergoing hip fracture surgery at two medical centers. Patients from the primary center were temporally allocated to a training cohort (n = 640) and an internal validation cohort (n = 160). An independent external validation cohort (n = 137) was collected from a second institution. The primary outcome was early PND, a composite of in-hospital delirium and cognitive decline assessed up to 3 months post-discharge. Least Absolute Shrinkage and Selection Operator (LASSO) regression was used for predictor selection. A multivariable logistic regression model was developed and visualized as a nomogram. Model performance was assessed via the area under the receiver operating characteristic curve (AUC), calibration plots, and decision curve analysis (DCA).

Results

Early PND occurred in 17.7% of the training cohort. Eight independent predictors were identified: age, body mass index, diabetes, frailty, ASA classification, albumin-to-fibrinogen ratio, neutrophil-to-lymphocyte ratio, and duration of surgery. The nomogram demonstrated good discrimination in the training cohort (AUC, 0.875; 95% CI, 0.841–0.910) and internal validation cohort (AUC, 0.869; 95% CI, 0.808–0.930). In the external validation cohort, the model maintained satisfactory discrimination (AUC, 0.775; 95% CI, 0.675–0.875). Calibration plots showed reasonable agreement between predicted and observed probabilities, and DCA indicated positive net clinical benefit across clinically relevant threshold probabilities in all cohorts.

Conclusions

We developed and externally validated a dynamic nomogram that accurately predicts early PND in older adults after hip fracture surgery. By incorporating readily available clinical and laboratory variables, this tool facilitates early risk stratification and may guide targeted perioperative interventions to improve cognitive outcomes.

Trial registration

Chinese Clinical Trial Registry, ChiCTR2500107395. Registered August 11, 2025.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12877-026-07125-2.

Introduction

Hip fracture constitutes a major public health problem, particularly in the context of global population aging. These patients are typically characterized by advanced age, frailty, and limited physiological reserves [1]. Consequently, the incidence of postoperative complications in this population is high [2]. Studies have reported that adverse events, such as pneumonia, cardiac complications, and acute renal failure, occur in approximately 3% to 15% of patients [3]. In contrast, PND has emerged as a particularly critical clinical concern, as its incidence is significantly elevated in hip fracture surgery compared to other surgical procedures, with reported rates ranging from approximately 13% to 50% [4, 5]. This complication is strongly associated with adverse clinical outcomes, including prolonged hospitalization, loss of functional independence, and elevated mortality [6, 7].

While previous studies have developed nomograms to predict postoperative delirium (POD) in older adults after hip fracture surgery [8–12], restricting risk assessment only to the acute in-hospital phase is insufficient. PND describes a broad spectrum of cognitive impairments diagnosed in the perioperative period [13]. This spectrum includes POD (occurring up to 7 days after surgery [14–16], delayed neurocognitive recovery (diagnosed up to 30 days postoperatively), and postoperative neurocognitive disorder (postoperative NCD, diagnosed from 30 days to 12 months after surgery). Cognitive decline that persists or emerges after discharge (e.g., within the first 3 months) is clinically critical, particularly for hip fracture patients. The initial three months following surgery constitute the critical window for functional recovery. Cognitive impairment during this phase can impair compliance with rehabilitation, increase the risk of falls, and reduce the ability to return to independent living. Existing prediction models, which often lack post-discharge follow-up, fail to identify patients at risk of these persistent early cognitive disorders that significantly burden both families and healthcare systems.

Prediction tools addressing this early post-discharge phase remain limited. The objective of this study was to develop and externally validate a dynamic nomogram for early PND. This outcome was defined as the occurrence of delirium or cognitive decline within three months following surgery. Consequently, a multicenter design was employed to construct a multivariable model based strictly on preoperative and intraoperative variables. The model was presented as both a static and a web-based dynamic nomogram to facilitate individualized risk assessment.

Methods

Study design and setting

This multicenter, retrospective cohort study was conducted at The Affiliated Hospital of Jiaxing University and Jiaxing Hospital of Traditional Chinese Medicine. The study was designed and reported in accordance with Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) statement [17]. The study was approved by the Ethics Committee of The Affiliated Hospital of Jiaxing University (2025-KY-460) and the institutional review board of the external center (JXTCM-IRB-2025–122). The committee granted a waiver of written informed consent due to the retrospective nature of the data analysis and the anonymization of all patient information.

Study population and cohort definition

We retrospectively enrolled patients aged ≥ 65 years who underwent surgery for hip fracture at two centers. The primary dataset was obtained from The Affiliated Hospital of Jiaxing University. Patients admitted between January 1, 2020, and December 31, 2024, were temporally allocated into two groups: the training cohort (January 1, 2020, to April 14, 2024; n = 640), which was used for model development, and the internal validation cohort (April 15, 2024, to December 31, 2024; n = 160). To evaluate the generalizability of the model, an independent external validation cohort was recruited from Jiaxing Hospital of Traditional Chinese Medicine, including 137 patients admitted between September 1, 2024, and August 15, 2025.

Patients were included if they met the following criteria: (1) Age ≥ 65 years; (2) Primary diagnosis of hip fracture (including femoral neck, intertrochanteric, or subtrochanteric fracture) confirmed by imaging; (3) Underwent surgical repair (internal fixation or arthroplasty) under general or neuraxial anesthesia; (4) Admitted within the specified study periods for each cohort.

Exclusion criteria: (1) Preexisting neurocognitive or psychiatric conditions: A documented diagnosis of dementia or another major neurocognitive disorder prior to admission, or chronic use of psychotropic medications (including benzodiazepines); (2) Specific comorbidities: A history of other major neurologic diseases known to affect cognitive function (e.g., Parkinson’s disease, epilepsy with recurrent seizures), or a diagnosis of chronic alcohol or substance use disorder that required medical treatment within the past year (3) Severe sensory impairment: Severe, uncorrected visual or auditory deficits that would preclude a valid delirium assessment; (4) Early postoperative mortality: Death occurring within 48 h after surgery; (5) Incomplete data: Absence of a primary outcome assessment or missing data for more than 20% of the predictor variables included in the final model.

Outcome assessment

The primary endpoint was early PND, defined as a composite outcome of cognitive impairment occurring from surgery through the first 3 months. This composite endpoint includes POD (acute phase), delayed neurocognitive recovery (1 month) and postoperative NCD (diagnosed at 3 months). First, in-hospital POD was identified retrospectively based on systematic electronic medical records. Our institutional protocol required daily delirium screening for all hip fracture patients from surgery until discharge. These assessments were performed by trained nursing and medical staff using the 3-Minute Diagnostic Confusion Assessment Method (3D-CAM) for ward patients and the Confusion Assessment Method for the ICU (CAM-ICU) for ICU patients [18, 19]. A positive diagnosis required the presence of acute onset and fluctuating course, inattention, and either disorganized thinking or an altered level of consciousness [20, 21].

Second, to evaluate cognitive trajectories after discharge, structured follow-ups were conducted at 1 month and 3 months. Cognitive status was assessed using a combination of face-to-face and telephone interviews. Patients returning to the outpatient clinic underwent the face-to-face Mini-Mental State Examination (MMSE). For patients unable to attend the clinic, the Telephone Mini-Mental State Examination (t-MMSE) was administered. The t-MMSE is a validated instrument that demonstrates high concordance with face-to-face testing in older populations [22, 23]. Significant cognitive decline was defined as a total score below 24 or a substantial decrease from the preoperative baseline [24]. Patients were defined as early PND if they met the criteria for in-hospital POD, delayed neurocognitive recovery (1 month) or postoperative NCD at 3 months.

Predictor variables and data collection

Candidate predictor variables were retrieved from the institutional electronic medical record and laboratory information systems. To ensure the model's applicability for early risk assessment, data collection was strictly limited to information available prior to the conclusion of the surgical procedure. The candidate variables included baseline demographics (age, sex, body mass index, educational level, smoking, and alcohol use) and specific comorbidities, including hypertension, diabetes, malignant tumors, prior stroke, heart failure, chronic obstructive pulmonary disease, and renal failure. The overall burden of comorbidity was quantified using the age-adjusted Charlson Comorbidity Index (aCCI), while physiological reserve was assessed using the 11-item modified Frailty Index (mFI-11). Preoperative laboratory data collected within 24 h of surgery included routine hematological and biochemical indices, as well as inflammatory markers such as C-reactive protein, interleukin-6, procalcitonin, and D-dimer. From these, composite biomarkers including the albumin-to-fibrinogen ratio (AFR) and neutrophil-to-lymphocyte ratio (NLR) were derived to quantify systemic inflammation and nutritional status. Intraoperative variables consisted of the type of surgery and anesthesia, operative duration, estimated blood loss, and intraoperative hemodynamic events. Detailed definitions and measurement units for all candidate variables are provided in Supplementary Table 1.

Sample size

The sample size was determined based on the Events Per Variable (EPV) principle recommended for multivariable prediction models. To ensure stable coefficient estimation and minimize overfitting, a minimum of 10 outcome events per predictor variable (EPV ≥ 10) is generally required. In our training cohort (N = 640), we observed 113 PND events (incidence rate 17.7%). Given that the final model incorporated 8 independent predictors, the resulting EPV was approximately 14.1, satisfying the methodological requirement. Furthermore, to handle the high dimensionality of the initial candidate variables and enhance model robustness, we employed LASSO regression as a regularization strategy for variable selection.

Statistical analysis

All statistical analyses were performed using SPSS version 26.0 (IBM SPSS Statistics, Armonk, NY, USA) and R version 4.0.2 (R Foundation for Statistical Computing, Vienna, Austria). A two-sided P value of less than 0.05 was considered statistically significant. Missing data for covariates were handled via multiple imputation by chained equations (MICE), with 20 imputed datasets generated to minimize bias.

The normality of quantitative data was assessed using the Kolmogorov–Smirnov test. Quantitative data with a normal distribution were expressed as mean ± standard deviation (SD) and compared using the independent samples t-test, while data with a skewed distribution were expressed as median [interquartile range (IQR)] and compared using the Mann–Whitney U test. Categorical data were presented as frequencies (percentages) and compared using the Chi-square test or Fisher’s exact test.

To address high dimensionality and mitigate overfitting, LASSO logistic regression was employed for predictor selection in the training cohort. Variables with non-zero coefficients at the optimal lambda (λ) value, determined by tenfold cross-validation, were selected. Multicollinearity among the selected predictors was ruled out by calculating the Variance Inflation Factor (VIF), with a threshold of < 5 indicating no severe collinearity. Furthermore, potential non-linear associations between continuous predictors and the log-odds of the outcome were evaluated using restricted cubic splines (RCS) to ensure the validity of linearity assumptions.

The selected variables were included in a final multivariable logistic regression model to identify independent predictors. Based on this model, a static nomogram and a web-based dynamic nomogram were constructed. The performance of the nomogram was comprehensively evaluated in the training cohort, followed by validation in the internal temporal validation cohort and the independent external validation cohort. Discrimination was assessed using the area under the receiver operating characteristic curve (AUC). Prediction accuracy was assessed by the calibration curves. The clinical utility of the nomogram was quantified using decision curve analysis (DCA). Finally, to address the potential heterogeneity of the composite outcome, sensitivity analyses were performed based on the temporal spectrum of neurocognitive impairment (in-hospital POD vs. post-discharge NCD) and the mode of assessment (face-to-face vs. telephone) to verify the model's robustness.

Results

Baseline characteristics of patients

The study flow and participant selection process are illustrated in Fig. 1. Initially, patients aged ≥ 65 years undergoing hip fracture surgery were screened for eligibility. Exclusion criteria were applied to ensure the reliability of cognitive assessment. A total of 55 patients were excluded for the following reasons: preexisting neurocognitive disorders (n = 13), severe visual or hearing deficits precluding assessment (n = 2), other major neurologic diseases (n = 3), early postoperative mortality within 48 h (n = 7), and critical missing data (n = 30).

Fig. 1.

Flowchart of participant selection and cohort allocation

Consequently, 937 eligible patients were included in the final analysis. Based on the study design and temporal/location criteria, participants were allocated into three cohorts: the training cohort (N = 640) and internal validation cohort (N = 160) from The Affiliated Hospital of Jiaxing University, and the independent external validation cohort (N = 137) from Jiaxing Hospital of Traditional Chinese Medicine. The demographic and clinical characteristics of the study population are detailed in Table 1. A total of 937 elderly patients undergoing hip fracture surgery were included in the final analysis. These participants were allocated to the training cohort (N = 640), the internal validation cohort (N = 160), and the independent external validation cohort (N = 137). As shown in Table 1, the baseline profiles appeared generally comparable across the three groups. We observed no statistically significant differences in key demographic variables, such as age (mean ± SD 75 ± 8 vs 75 ± 8 vs 76 ± 8 years; P = 0.508) and gender (P = 0.890). Similarly, the distributions of potential predictors, including frailty status (P = 0.217), ASA classification, and inflammatory biomarkers (e.g., NLR, P = 0.105; AFR, P = 0.841), were consistent among the cohorts. The incidence of the primary outcome (early PND) was 17.3% (111/640) in the training cohort, 18.8% (30/160) in the internal validation cohort, and 18.2% (25/137) in the external validation cohort (P = 0.903), suggesting a stable prevalence of the outcome across the study populations.

Table 1.

Patient demographics and baseline characteristics

Characteristic	Group			p-value
	Training set N = 640	Internal validation set N = 160	External validation set N = 137
Age,y	75 ± 8	75 ± 8	76 ± 8	0.508
Gender				0.890
Male	308 (48.1%)	80 (50.0%)	65 (47.4%)
Female	332 (51.9%)	80 (50.0%)	72 (52.6%)
BMI	22.6 ± 3.5	22.9 ± 3.7	22.4 ± 3.7	0.495
Educational level				0.508
≤ 6	444 (69.4%)	119 (74.4%)	102 (74.5%)
6 ~ 9	164 (25.6%)	36 (22.5%)	31 (22.6%)
≥ 9	32 (5.0%)	5 (3.1%)	4 (2.9%)
Smoking	129 (20.2%)	31 (19.4%)	26 (19.0%)	0.939
Drinking alcohol	115 (18.0%)	29 (18.1%)	26 (19.0%)	0.962
Hypertension	364 (56.9%)	88 (55.0%)	71 (51.8%)	0.544
Diabetes	169 (26.4%)	35 (21.9%)	34 (24.8%)	0.493
History of malignant tumors	14 (2.2%)	5 (3.1%)	7 (5.1%)	0.148
History of prior stroke	66 (10.3%)	20 (12.5%)	17 (12.4%)	0.620
History of heart failure	16 (2.5%)	4 (2.5%)	4 (2.9%)	0.950
History of COPD	12 (1.9%)	3 (1.9%)	3 (2.2%)	0.935
History of renal failure	8 (1.3%)	2 (1.3%)	0 (0.0%)	0.551
Frailty				0.217
Robust	302 (47.2%)	62 (38.8%)	58 (42.3%)
Prefrail	249 (38.9%)	77 (48.1%)	56 (40.9%)
Frail	89 (13.9%)	21 (13.1%)	23 (16.8%)
Charlson comorbidity index	5.00 (5.00, 6.00)	5.00 (5.00, 6.00)	5.00 (5.00, 6.00)	0.650
CRP	26 ± 44	19 ± 38	20 ± 41	0.105
Creatinine	74 ± 30	78 ± 46	77 ± 41	0.368
HCT	0.41 ± 0.04	0.41 ± 0.04	0.41 ± 0.03	0.203
Hb	111 ± 18	110 ± 19	109 ± 19	0.579
PLT	155 ± 58	155 ± 58	151 ± 49	0.731
PT	13.15 ± 0.77	13.05 ± 0.88	13.04 ± 0.88	0.176
APTT	35.3 ± 4.3	35.3 ± 4.7	35.3 ± 5.0	0.995
D-dimer	6065 (1670, 16,215)	5595 (1635, 16,170)	5120 (1590, 14,890)	0.884
Procalcitonin	0.17 (0.10, 0.35)	0.20 (0.11, 0.36)	0.20 (0.10, 0.36)	0.464
IL-6	92 (44, 138)	101 (55, 145)	95 (50, 145)	0.457
AFR	9.71 ± 1.13	9.71 ± 1.41	9.78 ± 1.51	0.841
NLR	3.45 ± 0.45	3.36 ± 0.56	3.41 ± 0.63	0.105
Blood loss volume,mL	135 ± 179	112 ± 118	115 ± 125	0.171
Duration of surgery, min	127 ± 46	123 ± 43	124 ± 40	0.666
Intraoperative hypotension	341 (53.3%)	82 (51.3%)	68 (49.6%)	0.703
Intraoperative bradycardia	207 (32.3%)	42 (26.3%)	37 (27.0%)	0.204
Intraoperative hypothermia	210 (32.8%)	55 (34.4%)	45 (32.8%)	0.930
Surgery type				0.885
Intramedullary nail	213 (33.3%)	57 (35.6%)	50 (36.5%)
Hemiarthroplasty	95 (14.8%)	24 (15.0%)	20 (14.6%)
Total hip replacement	224 (35.0%)	58 (36.3%)	49 (35.8%)
Cannulated screws	108 (16.9%)	21 (13.1%)	18 (13.1%)
Anesthesia type				0.764
General + regional	349 (54.5%)	77 (48.1%)	71 (51.8%)
General only	276 (43.1%)	78 (48.8%)	64 (46.7%)
Spinal ± sedation	15 (2.3%)	5 (3.1%)	2 (1.5%)
Early PND	111 (17.3%)	30 (18.8%)	25 (18.2%)	0.903

Abbreviations: BMI body mass index, COPD chronic obstructive pulmonary disease, IQR interquartile range, SD standard deviation

Lasso-­logistic method for variable screening and model fitting

A total of 45 candidate variables were evaluated by univariate analysis, and 34 variables with statistically significant associations were identified. These 34 variables were subsequently entered into the LASSO regression analysis to address high dimensionality. We used tenfold cross-validation to determine the optimal tuning parameter (λ). We applied the 1-standard-error rule (1-SE) to select the λ value, prioritizing a simpler model (Fig. 2). Based on this criterion, eight predictors with non-zero coefficients were retained.

Fig. 2.

Predictor selection using the Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression model. A LASSO coefficient profiles of the 45 candidate features. Each colored line represents the coefficient of a variable as it changes with the log(λ) sequence. B Selection of the optimal penalization coefficient lambda (λ) using tenfold cross-validation. The partial likelihood deviance (binomial deviance) is plotted against log(λ). The left vertical dotted line represents the minimum error (λ_min), and the right vertical dotted line represents the threshold for the 1-standard-error rule (λ_1se). We selected the λ_1se value to construct a parsimonious model, resulting in 8 predictors with non-zero coefficients

To further rule out multicollinearity among these selected predictors, we performed both correlation and variance inflation factor (VIF) analyses. The correlation matrix showed only weak to negligible correlations (Pearson’s coefficients ranging from −0.10 to 0.15). Furthermore, all VIF values ranged from 1.03 to 2.19, well below the conservative threshold of 5 (Supplementary Table 2 and Supplementary Fig. 2). Consequently, these eight independent predictors—Age, BMI, Diabetes, Frailty, ASA classification, AFR, NLR, and Duration of surgery—were included in the final multivariable logistic regression model to construct the nomogram.

Independent predictors and their non-linear associations with early PND

The eight variables selected by LASSO were entered into a multivariable logistic regression model. After adjusting for confounding factors, all eight variables remained independently associated with early PND (Table 2).

Table 2.

Multivariable logistic regression analysis of independent predictors for early PND

Variables	β	OR (95% CI)	p-value
Age	0.10	1.11 (1.07, 1.15)	< 0.001
BMI	−0.14	0.87 (0.79, 0.92)	< 0.001
Diabetes	0.72	2.05 (1.16, 3.60)	0.013
Frailty
Robust	Ref	-	-
Prefrail	0.72	2.06 (1.20, 3.55)	0.009
Frail	1.54	4.70 (2.41, 9.14)	< 0.001
ASA classification
Ⅱ	Ref	-	-
≥ Ⅲ	0.57	1.78 (1.10, 2.87)	0.019
AFR	−0.43	0.65 (0.53, 0.79)	< 0.001
NLR	0.58	1.79 (1.21, 2.64)	0.003
Duration of surgery, min	0.02	1.02 (1.01, 1.02)	< 0.001

Abbreviations: OR odds ratio, CI confidence interval, BMI body mass index, ASA American Society of Anesthesiologists, AFR albumin-to-fibrinogen ratio, NLR neutrophil-to-lymphocyte ratio, PND perioperative neurocognitive disorder

Regarding patient characteristics, advanced age was associated with increased odds of PND (OR 1.11; 95% CI 1.07–1.15; P < 0.001), while higher BMI was associated with decreased odds (OR 0.87; 95% CI 0.79–0.92; P < 0.001). Comorbidities and physiological status also showed significant associations: patients with diabetes had higher odds of PND (OR 2.05; 95% CI 1.16–3.60; P = 0.013), and those with an ASA classification ≥ III had an OR of 1.78 (95% CI 1.10–2.87; P = 0.019). Notably, frailty exhibited a graded association with the outcome; compared to the robust group, the odds of PND increased in the prefrail group (OR 2.06; 95% CI 1.20–3.55; P = 0.009) and were highest in the frail group (OR 4.70; 95% CI 2.41–9.14; P < 0.001).

Regarding biomarkers and surgical factors, a higher AFR was a protective factor (OR 0.65; 95% CI 0.53–0.79; P < 0.001), whereas a higher NLR was a risk factor (OR 1.79; 95% CI 1.21–2.64; P = 0.003). Prolonged duration of surgery was also associated with a slightly increased risk per minute (OR 1.02; 95% CI 1.01–1.02; P < 0.001).

To validate the linearity assumption of the continuous predictors in the logistic model, we performed restricted cubic spline (RCS) analysis (Supplementary Fig. 1). The analysis revealed no statistically significant non-linear associations between the log-odds of early PND and age (P_non-linear = 0.676), BMI (P_non-linear = 0.739), AFR (P_non-linear = 0.131), NLR (P_non-linear = 0.516), or duration of surgery (P_non-linear = 0.177). These results support the use of linear terms for these variables in the final prediction model.

Development and validation of a nomogram for PND prediction

Based on the multivariable logistic regression analysis, a static nomogram was constructed incorporating the eight independent predictors: Age, BMI, Diabetes, Frailty, ASA classification, AFR, NLR, and Duration of surgery (Fig. 3). Each predictor is assigned a score on the point scale, and the total score corresponds to a predicted probability of early PND. To facilitate clinical application, a dynamic web-based version of the nomogram was also developed and is available at https://wsr999815.shinyapps.io/dynnomapp/.

Fig. 3.

Diagnostic nomogram for predicting the probability of early-stage perioperative neurocognitive disorder (PND) in older adults with hip fracture. A Static Nomogram. To use the nomogram, locate the patient's value for each variable on the corresponding axis. Draw a vertical line upward to the "Points" scale to determine the score for that variable. Sum the scores for all eight variables to obtain the "Total Points". Finally, draw a vertical line downward from the total points to the "Risk" scale to estimate the individual probability of developing early-stage PND. (B) Dynamic Web-based Nomogram. A screenshot of the online calculator interface. The left panel allows clinicians to input specific patient characteristics. The right panel visually displays the predicted probability (point estimate) along with the 95% confidence interval. Abbreviations: BMI, body mass index; ASA, American Society of Anesthesiologists; AFR, albumin-to-fibrinogen ratio; NLR, neutrophil-to-lymphocyte ratio

The discrimination performance of the nomogram was evaluated using ROC analysis (Fig. 4a–c). In the training cohort, the model yielded an AUC of 0.875 (95% CI: 0.841–0.910). In the internal validation cohort, the AUC was 0.869 (95% CI: 0.808–0.930). In the independent external validation cohort, the AUC was 0.775 (95% CI: 0.675–0.875).

Fig. 4.

Evaluation of the nomogram's performance in the training, internal validation, and external validation cohorts. (A–C) Receiver operating characteristic (ROC) curves. The Area Under the Curve (AUC) values indicate the model's discrimination capability in the training (A), internal validation (B), and external validation (C) cohorts. (D–F) Calibration plots. The x-axis represents the predicted probability of early PND, and the y-axis represents the actual observed probability. The diagonal gray line represents perfect prediction by an ideal model. The solid black line represents the performance of the nomogram. (G–I) Decision curve analysis (DCA). The y-axis measures the net benefit. The red line represents the nomogram. The gray line represents the assumption that all patients have PND (Treat All), and the black horizontal line represents the assumption that no patients have PND (Treat None). The decision curves show the clinical utility of the model across different threshold probabilities in the training (G), internal validation (H), and external validation (I) cohorts

Calibration plots (Fig. 4d–f) were generated to assess the agreement between predicted probabilities and observed frequencies. The calibration curves for the training and internal validation cohorts demonstrated alignment with the ideal diagonal line. In the external validation cohort, the curve indicated a tendency towards overestimation of risk (Intercept = −0.448; Slope = 0.675).

DCA was performed to estimate clinical utility (Fig. 4g–i). The analysis indicated that using the nomogram to guide intervention strategies provided a greater net benefit than either the "treat-all" or "treat-none" schemes across the majority of threshold probabilities in the training and internal validation cohorts. In the external validation cohort, a positive net benefit was observed within a threshold probability range of approximately 5% to 40%.

Sensitivity analysis

To evaluate the potential heterogeneity introduced by the composite primary outcome, a sensitivity analysis was conducted by stratifying early PND into two distinct subtypes: POD (acute phase) and postoperative NCD (subacute phase). As detailed in Supplementary Table 3 and Supplementary Fig. 3, the incidence rates in the training cohort were 13.0% (83/640) for POD and 4.7% (30/640) for postoperative NCD.

Additionally, to evaluate the impact of the mixed-method follow-up approach on model performance, a sensitivity analysis was conducted by stratifying patients based on their mode of post-discharge cognitive assessment (face-to-face vs. telephone MMSE). The incidence of early PND was comparable between the 350 patients assessed face-to-face (17.1%; 60 events) and the 290 patients assessed via telephone (18.2%; 53 events). The nomogram demonstrated robust discrimination in the face-to-face group, yielding an AUC of 0.841 (95% CI: 0.780–0.903). While, as expected, performance was attenuated in the telephone-assessed group, the model maintained acceptable predictive accuracy with an AUC of 0.741 (95% CI: 0.670–0.812). These results, detailed in Supplementary Table 3 and visualized in Supplementary Fig. 4, suggest that the model is not unduly influenced by the assessment modality. This confirms that our hybrid approach was a reasonable strategy to minimize attrition bias while preserving the model's predictive validity.

Discussion

This study developed and externally validated a dynamic predictive nomogram for early PND in older adults with hip fracture. By analyzing data from multicenter cohorts, we identified eight independent predictors: age, BMI, diabetes, frailty, ASA classification, AFR, NLR, and duration of surgery. These findings are clinically significant because they provide a scientific basis for establishing a precise risk stratification tool. Unlike previous models that focused solely on in-hospital delirium, our model captures the critical "early recovery window" (up to 3 months), providing a generalizable application tool for clinical practice to identify high-risk patients who may suffer from persistent cognitive decline after discharge.

First, we observed that the incidence of early PND was approximately 17–18% across our training and validation cohorts. While risk factors for POD have been documented in prior studies [25], our results extend these findings by indicating that these predictors remain relevant throughout the early recovery period, including both acute delirium and subacute cognitive decline. The sensitivity analysis indicated consistent predictive performance for both the in-hospital POD subgroup and the post-discharge NCD subgroup. The incidence of POD (13.0%) was higher than that of postoperative NCD (4.7%), consistent with the expected clinical course of cognitive recovery. Despite the lower prevalence of the NCD outcome, the model demonstrated stable discrimination (AUC > 0.80) for both subtypes. This consistency suggests shared pathological pathways underlying both acute and subacute decline. Consequently, effective management requires extending risk assessment and monitoring into the post-discharge period.

We explored the biological predictors to provide evidence for targeted physiological optimization. Our findings indicated that high NLR and low AFR were potential predictors of early PND. This aligns with the neuroinflammation hypothesis, where peripheral inflammatory signals (amplified by surgical trauma and malnutrition) disrupt the blood–brain barrier and induce neuronal dysfunction [26]. Specifically, AFR reflects the interplay between nutritional reserve (albumin) and coagulation/inflammation (fibrinogen). Similarly, NLR serves as a composite marker of systemic stress, representing the disequilibrium between innate hyper-inflammation (neutrophils) and adaptive immune suppression (lymphocytes) [27]. Therefore, for older hip fracture patients with dysregulated inflammatory markers, we assume that medical professionals should not only focus on surgical repair but also implement perioperative immunonutrition strategies. Early correction of hypoproteinemia and administration of anti-inflammatory interventions could potentially dampen the systemic inflammatory response, thereby protecting the vulnerable aging brain.

Regarding diabetes, our findings (OR 2.05) align with a recent meta-analysis by Liu et al. [28], which confirmed a 44% increased risk of POCD in diabetic patients. The underlying mechanisms are likely multifaceted. First, chronic hyperglycemia accelerates the formation of advanced glycation end products (AGEs). The interaction of AGEs with their receptor (RAGE) activates the NF-κB pathway and the NLRP3 inflammasome, creating a pro-inflammatory microenvironment that amplifies the neuroinflammatory response triggered by surgical trauma [29, 30]. Second, central insulin resistance disrupts the PI3K/Akt signaling pathway and activates GSK-3β, leading to Tau hyperphosphorylation and synaptic plasticity impairment (LTP deficits), which diminishes the brain's cognitive reserve [31]. Furthermore, recent studies highlight the role of hippocampal neuronal ferroptosis and blood–brain barrier (BBB) disruption as critical pathways [32]. Moreover, as highlighted by Sándor et al. [33], diabetic patients exhibit significantly lower cerebral tissue oxygen saturation (rSO₂) and greater hemodynamic instability during surgery. This vulnerability may stem from diabetic cardiac autonomic neuropathy, which impairs the autoregulation of cerebral blood flow in response to surgical stress, leading to transient cerebral hypoperfusion and neuronal injury. Alternatively, we must acknowledge that the observed association might not be purely causal. In geriatric populations, diabetes may serve as a surrogate marker for cumulative physiological decline. This introduces the possibility of residual confounding or selection bias, where diabetic patients represent a sub-population with inherently lower resilience to surgical stress, independent of specific glycemic mechanisms.

Finally, we constructed a nomogram based on the results of the multivariable regression analysis. To further improve clinical convenience, we developed a web-based dynamic online nomogram. By simply inputting the 8 available clinical variables, clinicians can instantly calculate an individualized risk probability. The ROC curves, calibration plots, and DCA curves in both internal and external validation cohorts suggest that the model has acceptable discrimination, calibration, and clinical benefit. Notably, unlike many prior studies which lacked external validation, our study supported the model's generalizability in an independent hospital setting (Jiaxing Hospital of Traditional Chinese Medicine). Although a slight calibration drift was observed in the external cohort, the model maintained a positive net benefit within the clinically relevant threshold range (5%–40%), suggesting that this tool is satisfactory for screening high-risk patients across different clinical practices.

However, some limitations in our study should be acknowledged. First, due to the retrospective cohort design, data collection relied on electronic medical records. While we mitigated information bias by using systematic nursing flowsheets for delirium assessment, residual confounding from unmeasured variables—specifically social support, preoperative cognitive trajectory, and intraoperative EEG data—remains possible. Second, the definition of the follow-up window warrants caution. Our study focused on early PND (up to 3 months), whereas the standard PND spectrum extends to 12 months. This restriction means our model does not capture late-onset cognitive decline, potentially underestimating the total disease burden. Consequently, the predictive performance of our nomogram is strictly applicable to the early recovery phase, and our findings may not be directly comparable to studies employing longer observation periods. Third, heterogeneity in outcome measurement is a limitation. The study utilized different tools for different phases (3D-CAM/CAM-ICU for delirium vs. MMSE for cognitive decline) and employed a hybrid follow-up approach (face-to-face vs. telephone). Although our sensitivity analysis confirmed that the model remained robust across assessment modes, this inconsistency may introduce non-differential misclassification bias. Finally, while we performed external validation, future large-scale prospective studies are needed to further confirm the model's generalizability across diverse healthcare settings.

In the future, we recommend conducting prospective, multicenter, large-sample studies with longer follow-up periods (6–12 months) to further validate the trajectory of cognitive recovery. Future research should also integrate novel biomarkers and intraoperative monitoring data to refine the predictive precision of the nomogram.

Abbreviations

3D-CAM

3-Minute Diagnostic Confusion Assessment Method

aCCI

Age-adjusted Charlson Comorbidity Index

AFR

Albumin-to-fibrinogen ratio

ASA

American Society of Anesthesiologists

AUC

Area under the receiver operating characteristic curve

BMI

Body mass index

CAM-ICU

Confusion Assessment Method for the Intensive Care Unit

CI

Confidence interval

DCA

Decision curve analysis

EPV

Events per variable

LASSO

Least Absolute Shrinkage and Selection Operator

mFI-11

11-Item modified Frailty Index

MICE

Multiple imputation by chained equations

MMSE

Mini-Mental State Examination

NCD

Neurocognitive disorder

NLR

Neutrophil-to-lymphocyte ratio

OR

Odds ratio

PND

Perioperative neurocognitive disorder

POD

Postoperative delirium

RCS

Restricted cubic splines

SD

Standard deviation

t-MMSE

Telephone Mini-Mental State Examination

TRIPOD

Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis

VIF

Variance Inflation Factor

Authors’ contributions

Shirong Wei and Chaobo Ni performed the statistical analysis. Sitong Zhou and Tong Zhi drafted the manuscript. Xin Xiang, Junwen Tu and Zhangtian Xia were responsible for data collection. Tesheng Gao served as the Site Principal Investigator for the external validation center (Jiaxing Hospital of Traditional Chinese Medicine). Yungong Wang and Qihongshen provided supervision and guidance. Ming Yao and Huadong Ni designed the study and are the corresponding authors. All authors contributed to the article and approved the submitted version.

Funding

This experiment was partially funded by the Scientific Research Fund of National Health Commission-Zhejiang Provincial Health Major Science and Technology Plan Project (WKJ-ZJ-2448), Zhejiang Provincial Program of Traditional Chinese Medicine Science and Technology (2024ZL170), “Xingyao Nanhu” Leading Talent Program of Jiaxing City (2022-XYNHCXTD-001), the Clinical Key Specialty of Zhejiang Province Anesthesiology (2023ZJZK001) and the Zhejiang Province Compact yet High-performing Clinical Innovation Team (CXTD202502014).

Data availability

The data that support the findings of this study are not openly available. Deidentified individual participant data may be made available to qualified researchers for non-commercial purposes upon reasonable request to the corresponding author (Huadong Ni, at huadongni@zjxu.edu.cn), pending the submission of a sound research proposal and execution of a data sharing agreement.

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the Declaration of Helsinki (as revised in 2013. This study was approved by the Ethics Committee of The Affiliated Hospital of Jiaxing University (2025-KY-460) and Jiaxing Hospital of Traditional Chinese Medicine (JXTCM-IRB-2025–122). This study was conducted in accordance with the principles of the Declaration of Helsinki. The study protocol was reviewed and approved by the Ethics Committee of The Affiliated Hospital of Jiaxing University. The committee granted a waiver of the requirement for written informed consent from the participants due to the retrospective nature of the study, which involved the analysis of anonymized data from existing medical records and posed minimal risk to the patients.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.