The cohort was assembled from 418 surgical admission records for adults with imaging-confirmed acute femoral neck fractures who were scheduled for surgery. After exclusions for pathological fracture or high-energy multiple trauma (24 records), severe psychiatric disorder or long-term psychotropic medication use (15), severe cognitive impairment (65) and incomplete core clinical or questionnaire data (43), 271 admissions remained. Anxiety was measured with the anxiety subscale of the Hospital Anxiety and Depression Scale (HADS-A), administered 24 to 48 hours after admission and at least 12 hours before surgery by trained researchers in face-to-face bedside interviews. Using a prespecified cutoff of a HADS-A score greater than 8, 79 admissions (29.2 percent) were classified as having significant preoperative anxiety and 192 (70.8 percent) were not. Because the analytic file retained only the binary classification and not the underlying HADS-A totals or item responses, the researchers could not reclassify outcomes using alternative cutoffs or run score-based sensitivity analyses.

Baseline comparisons revealed a distinctive profile in the anxiety group. These admissions involved older patients, with a mean age of 70.39 years compared with 65.48 years in the non-anxiety group, and lower mean albumin concentrations of 39.25 versus 40.66 grams per litre. Preoperative pain scores on a visual analogue scale did not differ significantly between groups, averaging 3.81 versus 3.59. The psychological and social measures separated the groups more clearly: mean scores on the Connor-Davidson Resilience Scale (CD-RISC) were 48.66 in the anxiety group versus 64.36 without anxiety, scores on the Multidimensional Scale of Perceived Social Support (MSPSS) were 48.05 versus 57.80, and scores on the Surgical Fear Questionnaire (SFQ) were 11.47 versus 8.08, all with p values below 0.001. The frailty index was also higher in the anxiety group at 0.26 versus 0.20, and only 32.9 percent of anxious admissions had familiar faces present at admission, compared with 53.6 percent of the rest.

The variables were organised according to a biopsychosocial framework, in which physiological vulnerability was represented by frailty and comorbidity measures, psychological factors by resilience and surgical fear, and social context by perceived social support and the presence of familiar faces at admission. Clinical variables drawn from the electronic health record included age, gender, body mass index, haemoglobin, albumin, the Charlson comorbidity index, American Society of Anesthesiologists grade, injury-to-admission time, surgical and anaesthetic history and preoperative pain score. Administrative fields such as case order, operating room entry time and identifiers were excluded, as was the outcome itself, leaving 32 numeric candidate predictors. No outcome or predictor values were missing, so a median imputer retained in the pipeline did not actually replace any observations.

The validation design was deliberately conservative. Feature reduction used least absolute shrinkage and selection operator (LASSO) logistic regression, with the inverse-regularisation parameter tuned inside each inner training fold, and the entire selection procedure was repeated independently in each of five outer folds so that no single fixed feature set was imposed across the analysis. Each inner pipeline also refitted standardisation and the synthetic minority oversampling technique (SMOTE) to address class imbalance before model fitting. Outer validation used fivefold stratified group cross-validation with the patient identifier as the grouping variable, ensuring that records from the same patient never appeared in both training and test data. Within each outer training set, roughly 75 percent of records were used for fitting and 25 percent for Platt sigmoid calibration, and hyperparameters were searched over 18 random combinations per fold. The outer test fold was never used for preprocessing, feature selection, resampling, tuning or calibration.

Performance differences among the three algorithms were small and statistically indistinguishable. Random forest recorded the highest AUC point estimate at 0.932 (95 percent confidence interval 0.900 to 0.959), followed by LightGBM at 0.928 (0.894 to 0.957) and XGBoost at 0.921 (0.885 to 0.953). Paired bootstrap comparisons of AUC differences all included zero: random forest minus XGBoost was 0.011 (95 percent CI -0.011 to 0.033), LightGBM minus XGBoost was 0.007 (-0.013 to 0.029), and LightGBM minus random forest was -0.004 (-0.027 to 0.019). Brier scores ranged narrowly from 0.096 to 0.099. Random forest’s calibration intercept was 0.041 (95 percent CI -0.325 to 0.386) and its calibration slope 1.017 (0.690 to 1.571), and the corresponding XGBoost and LightGBM calibration statistics were similarly compatible with adequate calibration, though with wide intervals. The authors state that the overlapping intervals and paired comparisons do not support superiority of any one algorithm in this dataset.

Feature selection proved highly stable for the psychological and social measures. CD-RISC, MSPSS and SFQ were retained by LASSO in all five outer folds for all three models, while the frailty index was retained in three of five folds for each model. Preoperative pain score, despite showing no significant unadjusted difference between groups, was retained in three of five folds for XGBoost and LightGBM and two of five random-forest folds, which the researchers interpret as evidence that it sometimes added information conditional on other variables rather than as proof of an independent association. Out-of-fold TreeSHAP analysis on XGBoost identified CD-RISC, MSPSS, SFQ and the frailty index as the four leading contributors by mean absolute contribution. Higher resilience and social support values generally produced negative contributions, lowering the estimated probability of anxiety, whereas higher surgical fear values and, less consistently, higher frailty values produced positive contributions. The authors caution that SHAP values describe model behaviour, not causal effects or clinical validity.

An exploratory analysis of random-forest out-of-fold probabilities divided admissions into probability tertiles. Observed anxiety occurred in none of 91 admissions in the low tertile (0.0 percent, 95 percent CI 0.0 to 4.1 percent), in 13 of 90 admissions in the middle tertile (14.4 percent, 95 percent CI 8.6 to 23.2 percent), and in 66 of 90 admissions in the high tertile (73.3 percent, 95 percent CI 63.4 to 81.4 percent). The corresponding probability ranges were 0.0000 to 0.0435, 0.0438 to 0.3687 and 0.3688 to 1.0000. The researchers stress that these tertiles are sample-dependent descriptions of internal classification ordering and are not validated clinical cut-off points, and that decision curve analysis and tertile results were generated with 2000 patient-cluster bootstrap resamples to quantify uncertainty.

The study carries substantial limitations that the authors set out explicitly. It was a single-centre retrospective analysis with only 79 anxiety events, a limited event count that, together with class imbalance and fold-specific calibration, can leave residual overfitting despite nested validation. The source records did not preserve field-level acquisition times for every candidate predictor, and HADS-A and several psychosocial measures were collected during the same preoperative assessment phase, so the models cannot establish that predictors would be available at a future decision point. Dataset shift, subgroup calibration and algorithmic bias were not evaluated, and the models have not been tested externally or prospectively. The findings therefore do not establish clinical utility, nursing workflow benefit or patient benefit, and the researchers recommend that future studies use prospectively collected data with a prespecified predictor-availability schedule, retain HADS-A totals and item responses, externally validate calibration and discrimination, assess subgroup performance and algorithmic bias, and test whether model-informed structured nursing assessment changes patient-centred outcomes. Until such work is done, the authors conclude, structured assessment should continue to rely on validated anxiety screening and clinical judgement.

Subject of Research: Explainable machine learning classification of preoperative anxiety in femoral neck fracture patients

Article Title: Explainable Machine Learning Models for Classifying Preoperative Anxiety in Patients With Femoral Neck Fractures: A Retrospective Cohort Study

Article References: Zhang, X., Zheng, S., Zhang, Z., Yang, P., Wei, Q., Jin, M., & He, F. (2026). Explainable Machine Learning Models for Classifying Preoperative Anxiety in Patients With Femoral Neck Fractures: A Retrospective Cohort Study. Nursing Open, 13(10), Article e70919. https://doi.org/10.1002/nop2.70919

Image Credits: AI Generated

DOI: 10.1002/nop2.70919

Keywords: machine learning, preoperative anxiety, femoral neck fracture, XGBoost, random forest, LightGBM, SHAP, nested cross-validation, HADS-A, biopsychosocial model, frailty, nursing