Development and validation of a nomogram model for prolonged length of stay in spinal fusion patients: a retrospective analysis

Abstract

Objective

To develop a nomogram model for the prediction of the risk of prolonged length of hospital stay (LOS) in spinal fusion patients.

Methods

A retrospective cohort study was carried out on 6272 patients who had undergone spinal fusion surgery. Least absolute shrinkage and selection operator (LASSO) regression was performed on the training sets to screen variables, and the importance of independent variables was ranked via random forest. In addition, various independent variables were used in the construction of models 1 and 2. A receiver operating characteristic curve was used to evaluate the models’ predictive performance. We employed Delong tests to compare the area under the curve (AUC) of the different models. Assessment of the models’ capability to improve classification efficiency was achieved using continuous net reclassification improvement (NRI) and integrated discrimination improvement (IDI). The Hosmer–Lemeshow method and calibration curve was utilised to assess the calibration degree, and decision curve to evaluate its clinical practicality. A bootstrap technique that involved 10 cross-validations and was performed 10,000 times was used to conduct internal and external validation. The were outcomes of the model exhibited in a nomogram graphics. The developed nomogram was validated both internally and externally.

Results

Model 1 was identified as the optimal model. The risk factors for prolonged LOS comprised blood transfusion, operation type, use of tranexamic acid (TXA), diabetes, electrolyte disturbance, body mass index (BMI), surgical procedure performed, the number of preoperative diagnoses and operative time. The diagnostic performance of the nomogram model was satisfactory, with AUC values of 0.784 and 0.795 for the internal and external validation sets, respectively. Model discrimination was favourable in both the internal (C-statistic, 0.811) and external (C-statistic, 0.814) validation sets. Calibration curve and Hosmer-Lemeshow test showed acceptable agreement between predicted and actual results. The decision curve shows that the model provides net clinical benefit within a certain decision threshold range.

Conclusions

This study developed and validated a nomogram to identify the risk of prolonged LOS in spinal fusion patients, which may help clinicians to identify high-risk groups at an early stage. Predictors identified included blood transfusion, operation type, use of TXA, diabetes, electrolyte disturbance, BMI, surgical procedure performed, number of preoperative diagnoses and operative time.

Introduction

Spinal fusion surgery refers to a crucial surgical technique for the treatment of spinal trauma and degenerative ailments. This procedure is mainly aimed at the stabilisation of the spine along with simultaneous relief of pain, promotion of functional recovery, restoration of spinal physiological curvature, correction of deformities and improvement of the quality of life by decompressing the spinal cord and nerves. Spinal fusion surgery has become the prevailing surgical technique for various spinal ailments, with degenerative disorders being the most frequent indications [1, 2]. In the United States, the use of spinal fusion surgery increased considerably. This phenomenon has been attributed to an improved understanding of the biological aspects of the spine, enhanced diagnostic imaging techniques, advances in spinal fusion technology and an overall increase in life expectancy [3]. Between 1996 and 2008, the number of patients who underwent spinal fusion surgery increased by 2.4 times (137%), from 174,223 to 413,171 annually (p < 0.001). This increased rate was significantly higher than that of other surgeries [4]. Our hospital saw an around 2.2 times increase in the number of patients who underwent spinal fusion surgery between 2017 and 2022. Spinal fusion surgery is an expensive operation, with the average hospital bill exceeding $34,000, excluding professional fees [1]. The current managed care landscape, which is characterised by reduced reimbursements and escalating expenses, necessitates effective cost control. Strengthening of management and regulation of the duration of hospital stay are cost-saving strategies that conserve valuable human resources.

Following the introduction of disease diagnosis groups (DRG) in China, medical insurance providers started basing the reimbursement for spinal fusion surgery on a fixed cost for each admission in the DRG system. As a result, hospitals currently focus on the reasonable control of the length of hospital stay (LOS) and reduction of other hospitalisation costs. Prolonged hospital stays can increase the probability of contracting nosocomial infections, which can result in increased complications and mortality rates [5, 6]. Therefore, the early prediction of a patient’s risk of prolonged LOS is important in offering guidance in treatment decisions and ensuring cost effectiveness while providing appropriate care. Although current research has mainly concentrated on the identification of factors that influence prolonged LOS, given the increased number of spinal fusion surgeries being performed and the paucity of assessment tools for prolonged LOS for spinal fusion patients, coupled with the shortage of nurses and implementation of the cost-control policies of DRG in Medicare, early and accurate prediction of hospital stay is urgently needed to guarantee cost-effective delivery of health care services. In this study, we investigated the risk factors that contribute to extended hospitalisation following spinal fusion surgery. We developed and compared multiple prediction models and identified the optimal one, which we used in the rapid and accurate assessment of the risk of prolonged hospitalisation of spinal fusion surgery patients. The developed model also enables early identification and management of high-risk individuals. Thus, this approach may improve patient outcomes and provide a framework for cost-effective clinical management.

Materials and methods

Ethics statement and patient selection

The study protocol received approval from the Ethics Committee of Liuzhou Worker’s Hospital (ethics number: KY2024569) and adhered to the Declaration of Helsinki’s (2013) guidelines. All patients exempt from written informed consent. This study involved a retrospective analysis that included all patients who underwent spinal fusion surgery at the Liuzhou Worker’s Hospital from January 1, 2020, to March 31, 2023. The inclusion criteria were as follows: (1) all surgical procedures comprising spinal fusions; (2) complete case data and surgical records; (3) no contraindications to surgical treatment. The exclusion criteria included the following: (1) sudden disease aggravation; (2) history of spinal radiotherapy; (3) incomplete or illogical clinical data (e.g. negative age or duration of surgery or LOS). All data used in this work were sourced from the electronic medical record system of hospital. During the study period, the initial search using the International Classification of Surgery (ICD-9-CM-3 code 81.0) identified 5,120 patients who underwent spinal fusion surgery. The ICD-9-CM-3 code were as follows: 81.00, spinal fusion not otherwise specified; 81.01, atlas-axis spinal fusion; 81.02, other cervical fusion of the anterior column under the anterior technique; 81.03, other cervical fusion of the posterior column under the posterior technique; 81.04, dorsal and dorsolumbar fusion of the anterior column under the anterior technique; 81.05, dorsal and dorsolumbar fusion under the posterior technique; 81.06, lumbar and lumbosacral fusion in the anterior column under the anterior technique; 81.07, lumbar and lumbosacral fusion of the posterior column under the posterior technique; 81.08, lumbar and lumbosacral fusion of the anterior column under the posterior technique. Two investigators extracted and entered the data, and a third made the final decision in the case of conflicting information. All data collectors received training for data assessment and collection. Finally, the study included 5,036 patients. Based on the principle of randomisation, 70% of the participants were randomly assigned to the training data set, and the remaining 30% were assigned to the internal validation data set. The training and internal validation data sets consisted of 3,525 and the 1,511 cases, respectively. To conduct the external validation, 1,236 patients who underwent spinal fusion surgery at the Liuzhou People’s Hospital between 1 January 2020 and 30 September 2024 were used as the external validation set after meeting the inclusion and exclusion conditions described above.

Definition

The LOS referred to the actual number of days spent by a patient in the hospital, and it was calculated as the difference between the dates of admission and discharge [7]. The number of days in hospital involving admission and discharge were on the same day was calculated as 1 day. With reference to the relevant literature [8,9,10], the cut-off point for prolonged LOS, which was defined as exceeding the cut-off point, was 75% of the LOS of spinal fusion patients. We used separate cut points for each sets in order to control for the inherent differences in LOS between these different sets.

Clinical information

The electronic medical record system 3.5 (ZOE SOFT, China) was used in the collection of baseline data, preoperative and intraoperative conditions, medication use, and comorbidities of subjects. The following variables were considered in this study: age, gender, height, weight, number of preoperative diagnoses, the number of hospitalisations, ethnicity, marital status, occupation, education level, surgical history, critical patients (which means that the patients themselves were critically ill), surgeon’s professional title, blood transfusions, admission route, surgical type, blood type, use of TXA,American Society of Anesthesiologists (ASA) score, the week and season of surgery, fusion segments, surgical procedures and presence of co-morbidities. The included co-morbidities comprised diabetes, sleep apnoea, chronic obstructive pulmonary disease (COPD), cardiac arrhythmia, heart failure, coronary atherosclerosis, end-stage kidney disease or chronic kidney disease (ESKD/CKD), anxiety or depression, hypertension, peripheral vascular disease, anaemia, coagulation disorders, malignancy, electrolyte disorders and osteoporosis.

Statistical analysis

All statistical analyses were accomplished on R 4.2.2 software (R Foundation for Statistical Computing, Vienna, Austria) and involved the usage of several packages, including ‘tableone’, ‘rms’, ‘glmnet’, ‘caret’, ‘randomForest’, ‘readr’, ‘pROC’, ‘rmda’, ‘nricens’ and ‘PredictABEL’ .Continuous data that conformed to normal distribution were presented as the mean (± standard deviation), i.e. (‾X ± S), and the difference between the two groups was analysed via t-test. Data that lacked normal distribution were presented as the median (percentile), i.e., M (P₂₅–P₇₅), and the difference between the two groups was analysed through the rank-sum test. Categorical data were expressed as proportions and percentages and analysed via Chi-squared test (or Fisher’s exact probability method) for the comparison of distributions between the two groups. Predictors were screened using the least absolute shrinkage and selection operator (LASSO) regression and random forest methods, and the results were used to identify two distinct independent variables and create a model for comparison. Comparison of the models was conducted based on their receiver operating characteristic (ROC) curves, calibration curves and and decision curve analysis (DCA). To evaluate the model differentiation, we calculated the area under the curve (AUC). The differences between the AUCs of the models were subjected to Delong test, and net reclassification improvement (NRI) and integrated discrimination improvement (IDI) were usedto compare the efficacy of various models in terms of the improvement in classification efficiency. NRI > 0 and IDI > 0 indicate a positive improvement [11]. The degree of calibration was evaluated through the Hosmer-Lemeshow statistical test for goodness of fit. The optimal model was selected, and the nomogram was drawn. For the internal and external velidation of the nomogram model, the bootstrap method with 10,000 repetitions of sampling was applied. Two-sided p-values less than 0.05 were considered statistically significant.

Results

Characteristics of the subjects

The patient selection methodology is shown in Fig. 1. A total of 6,272 subjects were included in this study, including 2,868 males (45.73%) and 3,404 females (54.27%). The mean duration of surgery was (2.45 ± 1.23) hours. A total of 3,525 subjects were included in the training set, with 2,693 in the normal LOS group and 832 in the prolonged LOS group. A total of 1,511 cases were included in the internal validation set, with 1,163 cases in the normal LOS group and 348 cases in the prolonged LOS group. A total of 1,236 cases were included in the external validation set, including 968 cases in the normal LOS group and 268 cases in the prolonged LOS group. Table 1 summarises the clinical characteristics of the training set. The two groups exhibited statistically significant differences in terms of the age distribution, the number of preoperative diagnoses, the number of hospitalisations, operative time, BMI, marital status, occupation, education, surgical history, critical patients, blood transfusion, admission route, type of surgery, use of TXA, ASA  ≥ 3, diabetes, heart failure, coronary artery disease, ESKD/CKD, anxiety or depression, hypertension, peripheral vascular disease, anaemia, malignant neoplasm, electrolyte disturbance, the type of surgical procedure, fused bone segment and week of surgery. However, the other characteristics revealed no statistically significant differences (p > 0.05).

Table 1 Clinical characteristics of the subjects in the training set

Distribution of LOS among study participants

Figure 2 shows the distribution of LOS for all patients in the training set. The median and interquartile range of LOS for patients who underwent spinal fusion surgery in this derivation cohort was 10 [7.00, 14.00] days, with a range of 3 to 216 days. Therefore, those with LOS exceeding 14 days were considered to belong to the prolonged LOS group. The normal LOS group had an average stay of (9.02 ± 2.52) days, while the prolonged LOS group had an average stay of (27.45 ± 21.28) days. The median LOS for the normal and prolonged groups was 9.00 [7.00, 11.00] days and 21.00 [17.00, 28.00] days, respectively.

Screening models

LASSO regression screening variables

LASSO regression analysis was conducted on all indicators in the training set to achieve a high degree of fit and only the most essential variables were incorporated in the prediction model. The process of variable selection is shown in Fig. 3A, demonstrating the contraction of variables as the penalty term lambda increases. During the cross-validation process of the LASSO regression, we used the mean squared error (MSE) as the evaluation index. The value associated with the minimum MSE(Lambda. min = 0.005025126, log=-5.2983), at which 31 variables were selected. Further increasing lambda to the threshold of one standard error (SE) above the minimum MSE (Lambda.1se = 0.025125634, log=-3.689) retained 9 variables, as shown in Fig. 3B. Due to the large number chosen for Lambda.min, we selected Lambda.1se and found that the coefficients of these variables were not zero, indicating their significant contribution to the predictive results of the model. The following predictors were screened: blood transfusion, operation type, TXA use, diabetes, electrolyte disturbance, BMI, surgical procedure, the number of preoperative diagnoses and operation time. Based on the findings of LASSO regression analysis, these predictors were applied in the construction of a multi-factor logistic regression model (Model 1).

Random forest screening variables

The random forest machine learning method was used to screen all variables in the training set (Fig. 4A and B). Based on the Gini coefficient, six independent variables with importance scores > 70 were used as independent variables for subsequent analysis. The variables identified for inclusion were operative time, number of preoperative diagnoses, age, surgical procedure performed, week day of surgery, and occupation. The resulting multi-factor logistic regression model was defined as Model 2.

Fig. 1

Flow chart of research objects

Fig. 2

Distribution of LOS among the study participants in training set

Fig. 3

LASSO regression model for clinical feature selection. (A) Plot of the model’s coefficient distribution for logarithmic (lambda) sequences at different penalty levels. (B)For cross verification, the first column represents the minimum error, while the second column represents the cross-verification error of 1 standard deviation

Fig. 4

Random forest for selecting clinical features. (A)The relationship between tree and error in the random forest. (B) Random forest variable importance ranking

Test and evaluation of prediction model

The training set for Model 1 had a prediction probability AUC of 0.775 (95% confidence interval (CI): 0.756–0.793), whereas the Model 2 set attained a value of 0.761 (95% CI:0.742–0.780), Table 2 shows the performance index values for each model. Based on the findings displayed in Fig. 5, both Model 1 and Model 2 possess the ability to distinguish prolonged LOS in spinal fusion patients. During the comparison of the AUCs of Models 1 and 2, the Delong test results indicated the higher AUC of Model 1 than Model 2 in the training set, with statistical significance (Z = 2.451, p = 0.0014). NRI and IDI were also used to evaluate the capability of the two models to improve the classification effect in the training set. The results showed that in the training set, compared with Model 2, the prediction efficiency of Model 1 was positively improved, NRI was 2.22% (95%CI: 1.22–19.21%), IDI was 1.31% (95%CI: 0.19–2.43%), and the differences were statistically significant (p < 0.05).

Table 2 Predictive power of models 1 and 2 in training and validation set

Fig. 5

In training set, ROC curve for predicting prolonged LOS in spinal fusion patients

Fig. 6

In training set, calibration curve of model for predicting prolonged LOS in spinal fusion patients

Fig. 7

In training set, decision curve analysis of model predicting prolonged hospitalization in spinal fusion patients

Accuracy evaluation of the models

In the training set, the Hosmer-Lemeshow goodness-of-fit test indicated that Model 1 and Model 2 were a good fit for risk of prolonged LOS among spinal fusion patients, with χ² values of 9.599 and 5.534, and p-values of 0.294 and 0.699, respectively. The calibration curve in Fig. 6 also suggested that Models 1 and 2 have good calibration capability.

Clinical applicability

Decision curve analysis (DCA) was used to evaluate the effectiveness of the prediction model. The DCA results suggested that if the patient and physician threshold probabilities were 2-61% for Model 1 and 3-60% for Model 2, then using the model to predict the risk of prolonged LOS in patients undergoing spinal fusion surgery would provide more benefit than intervention or no intervention for all patients. The net benefit of Model 1 is greater than that of Model 2 (Fig. 7).

Optimal model and model visualization

The minimum AIC principle was used to compare Model 1 and Model 2, and the difference was statistically significant (AIC: 3183.924 vs. 3273.798, χ² = 79.875, p < 0.0001). The final prediction model is defined as Model 1. The established nomogram of Model 1 was used to predict the probability of prolonged LOS in patients undergoing spinal fusion surgery (Fig. 8). The variables appeared in rows two to ten, in which the values were acquired from the patient. The first row contained the point assigned to each variable’s measurement. The assigned points for all variables were then summed, with the total located on the line of total points. Subsequently, the total points were identified, and the probability of prolonged LOS was determined by drawing down vertical line to the bottom line. The nomogram displayed the scores for blood transfusion, emergency surgery, absence of TXA, diabetes, electrolyte disturbance, and BMI ≥ 24 kg/m²: 43, 24, 24, 36, 39 and 29 points, respectively. The points for the spinal fusion varied depending on the method used, such as atlas-axis spinal fusion, other cervical fusion of the anterior column under the anterior technique, other cervical fusion of the posterior column under the posterior technique, dorsal and dorsolumbar fusion of the anterior column under the anterior technique, dorsal and dorsolumbar fusion under the posterior technique, lumbar and lumbosacral fusion in the anterior column under the anterior technique, lumbar and lumbosacral fusion of the posterior column under the posterior technique, lumbar and lumbosacral fusion of the anterior column under the posterior technique. These methods scored 19, 9, 25, 45, 14, 2, 0 and 3 points, respectively. The scores increased to six points for every two additional preoperative diagnoses. With every extra hour of surgery, the score increased by approximately 8 points. The sum of scores of various risk factors corresponded to the probability of prolonged LOS in patients with spinal fusion. For example, spine fusion surgery patients with BMI ≥ 24 kg/m² underwent elective surgery, surgical procedure was 81.02, blood transfusion, use of TXA, operation time of 4.86 h, and nineteen preoperative diagnoses. The patients did have electrolyte disorders and diabetes, with corresponding scores reaching 29, 13, 9, 43, 16, 44.5, 65.5, 39 and 36 points, respectively. The total score of 295 indicates the estimated risk of prolonged LOS (0.989). The nomogram model achieved a C-index values of 0.767 (95% CI:0.729–0.805), indicating good agreement between predicted and actual results.

Fig. 8

Nomogram for predicting prolonged LOS in spinal fusion patients

Fig. 9

Internal validation set, apparent performance of the nomogram for risk of prolonged LOS in spinal fusion patients. (A) ROC curve of risk prediction nomogram; (B) Calibration curve of risk prediction nomogram; (C) Decision curve analysis of risk prediction nomogram

Fig. 10

External validation set, apparent performance of the nomogram for risk of prolonged LOS in spinal fusion patients. (A) ROC curve of risk prediction nomogram; (B) Calibration curve of risk prediction nomogram; (C) Decision curve analysis of risk prediction nomogram

Validation of the risk prediction model

Model performance and internal validation

When applied to an internal validation set, we developed a nomogram model with a sensitivity of 63.2%, specificity of 79.0%, positive predictive value (PPV) of 47.6% and negative predictive value (NPV) of 87.8% for predicting prolonged LOS (Table 2). The area under the receiver operating characteristic curve (AUROC) was 0.784 (95% CI:0.756–0.811), indicating good accuracy (Fig. 9A). Using the bootstrap method with 10,000 samples, the internal validation C-index was 0.811 (95% CI: 0.789–0.833), showing strong discrimination and predictive ability. The Hosmer-Lemeshow test was not significant (χ² = 8.806, p = 0.3589) and the calibration curve confirmed that the predictions of the nomogram were in good agreement with the actual results (Fig. 9B). The decision curve showed that this nomogram provided a net benefit in predicting prolonged LOS when the threshold probability was between 5% and 82% (Fig. 9 C).

Model performance and external validation

For all patients in the external validation set, the ROC curve for the model, with an AUC of 0.795(sensitivity, 0.627; specificity,0.831) (Table 2) and the C-statistic value of the developed model was 0.814 (95% CI:0.792–0.836), which showed good discrimination (Fig. 10A). Our model was also relatively well calibrated, and the Hosmer-Lemeshow goodness-of-fit value was 3.727 (p = 0.8809) in the external validation set (Fig. 10B). The DCA of the established model indicated positive net benefits with the threshold risk range of 5–93% (Fig. 10 C).

Discussion

Given the advancement of China’s medical system reform, the society is shifting its focus towards the development of a high-quality and efficient healthcare service system with Chinese characteristics. Achievement of this goal involves maintaining people’s health at the centre and promoting public welfare in medical and health undertakings. In addition, it requires the efficient utilisation of limited medical resources and addressing the issue of ‘difficulty in accessing medical treatment due to high costs’. LOS represents a crucial index of a hospital’s work efficiency, diagnosis and treatment standards, management level and nursing quality. Its effect on the medical quality, payment from medical insurance, hospitalisation costs and hospital benefit has become an crucial issue for hospital administrators, medical staff, patients and their families. Ensuring efficient, quality medical treatment requires effective prevention of extended hospitalization days. This approach will not only reduce patients’ hospitalization costs and their economic burden, but also increase hospital capacity and improve the utilization rate of beds. Ultimately, such a strategy will result in the improved social and economic benefits of hospitals overall. In addition, the accurate prediction of prolonged LOS and the implementation of active measures based on the predicted outcomes contribute to the improvement of surgical efficiency and reduction of medical costs. This method effectively improves hospital resource allocation at the management level. Based on common clinical indicators, a nomogram model was developed and validated this study, and it was used to predict extended hospital stays for patients undergoing spinal fusion surgery. The model can assist decision makers in the development of treatment strategies and cost-containment measures and provide a valuable reference for clinical practice.

In the study, Models 1 and 2 were constructed via LASSO regression and random forest screening, respectively, and comprehensively compared in terms of differentiation, calibration and clinical applicability. Optimal Model 1 was selected through comparison of the Delong test with INI and IDI and AIC. The included risk factors comprised blood transfusion, type of surgery, TXA use, diabetes, electrolyte disturbance, BMI, the number of preoperative diagnoses and operation time. Blood transfusion represents an independent risk factor for the LOS of spinal fusion patients. As presented in the nomogram, patients with blood transfusions is approximately 43 points. The receipt of blood transfusion receipt during treatment implies severe illness of spinal fusion patients. Husted et al. [12] associated the need for blood transfusions with prolonged LOS, as observed through the analysis of clinical data on 712 patients undergoing joint replacement surgery. This work reinforced and broadened this finding by revealing that such methods applies not only to joint replacement patients but also to those undergoing spinal fusions. Therefore, effectively reduction of the LOS to some extent can be attained thorough preoperative evaluation, minimisation of intraoperative bleeding and mitigation of patient transfusion risk.

TXA use during spinal fusion surgery is a protective factor of prolonged LOS, and patients who did not use TXA received a weight of approximately 24 points in the nomogram. As an antifibrinolytic medication, TXA binds to the lysine-binding site of plasminogen, which inhibits fibrinolysis and promotes haemostasis [13, 14]. TXA can reduce perioperative blood loss and the need for blood transfusions in patients with fractures [15, 16]. The protective effect of TXA may be associated with low blood loss and transfusion rates. Nonetheless, theoretically, TXA leads to delayed thrombosis and possible thrombotic events. The administration of TXA during spinal surgery may result in complications, including thrombosis, cerebral infarction or acute kidney failure [17, 18]. The potential complications associated with the use of TXA have been scrutinized clinically. Whether the use of TXA can decrease the likelihood of extended hospital stays and whether any associated risks exist must be considered. Ultimately, physicians should decide on the usage of TXA in clinical practice based on specific diagnosis and treatment plan.

According to Ibanez et al. [19], patients with preoperative mild renal insufficiency (glomerular filtration rate ≤ 60 mL/min/1.73 m²) and undergoing heart valve surgery experienced an increase in blood transfusion volume and prolonged LOS. Samer et al. [20] demonstrated that the occurrences of hypertension, diabetes and renal disease in emergency general surgery patients strongly predicted prolonged LOS. This study was conducted to examined various complications, including diabetes, arrhythmia, heart failure, coronary arteriosclerosis, ESKD, CKD, depression, hypertension, peripheral vascular disease, anemia, coagulation dysfunction, electrolyte disorders, and osteoporosis. Moreover, patients with diabetes and electrolyte disorders scored achieved 36 and 39 points on the nomogram, respectively. Physicians and healthcare providers in clinical practice should focus on the optimisation of patients’ electrolyte and blood glucose status prior to spinal fusion surgery and remain constantly aware of the risk of prolonged LOS resulting from disease exacerbation.

The identified risk factors for prolonged LOS comprised undergoing emergency surgery and the number of preoperative diagnoses. The patients who underwent emergency surgery nomogram approximately obtained 24 points. In addition, each increase in the two diagnoses prior to surgery increased the predictive model score by approximately 6 points. Possibly, patients requiring emergency surgery or suffering from multiple diseases exhibit more severe illness and extended their admission stay for diagnosis and treatment. This study and previous published research [21, 22] indicate that prolonged operative time is a factor in the prolonged LOS of patients, and each 1 h increase in the operative time increases the score of the predictive model by approximately 8 points. This may, however, depend on the surgeon’s skill and complexity of the procedure. Gardner et al. [23] revealed the association of surgical approach with prolonged LOS of patients undergoing radical prostatectomy, similar to the findings of this study. In addition, BMI ≥ 24 kg/m² is an independent risk factor for prolonged LOS. BMI ≥ 24 kg/m² has a score of 29 points in the nomogram. The decrease in the levels of elastin, increase in those of collagen V and VI and decrease in capillary density in obese individuals, accompanied with relative tissue hypoxia, impaired angiogenesis caused by chronic low-grade inflammation and micronutrient deficiency, negatively influence wound healing and show an association with wound infection in many surgical procedures [24, 25]. Therefore, overweight individuals are likely to have comorbidities or short-term complications, long recovery period, challenging anatomical conditions or longer surgical period. A study of 35,000 orthopaedic surgeries revealed the hospital stays were 1.1 days longer in morbidly obese patients [26]. In light of this finding, this outcome should inspire obese people to lose weight, and clinics should collaborate with hospital managers to conduct health education lectures on the prevention of obesity to avoid the prolonged LOS caused by obesity. Interestingly, in Model 2, increased the risk of prolonged LOS in spinal fusion patients, which reminded me the operative time, number of preoperative diagnoses, age, surgical procedure performed, week day of surgery, and occupation. A day-off week effect was also observed on the prolonged LOS associated with the day of the week surgery was performed. However, Friday surgery was considered a risk factor for prolonged LOS. This day is a neglected time point, and the findings indicate the ‘Friday effect’. This difference may be explained by the decreased number of medical staff working on weekends.

Strengths and limitations

In our research, all variables included in the predictive model are quantifiable predictors that are readily available to clinicians. Furthermore, the novel nomogram provides a visual scoring system for estimating the likelihood of prolonged LOS in patients with spinal fusion and performed well in the internal and external validation sets. The prediction model objectively calculates the probability of prolonged LOS in spinal fusion patients, helps clinicians identify and manage high-risk groups early, and provides a reference for evaluating prolonged LOS in spinal fusion patients. Early intervention in high-risk groups can reduce the risk of prolonged LOS.

However, there are limitations to this study. This is a retrospective study, and the evidence presented is not sufficient to prove causality. It is possible that some unidentified confounding factors were not taken into account in the analysis. But all the data analysed were traceable key characteristics from the hospital’s medical record system, which reduces the risk of potential errors. In addition, this study only included variables derived from electronic medical records. However, certain variables that have an impact on outcomes were not fully included. In the future, it would be useful to include additional laboratory variables, particularly newly derived biomarkers. Examples of such variables include white blood cell count, red cell to lymphocyte ratio (RLR) [27], neutrophil to lymphocyte ratio (NLR) [28], and monocyte to lymphocyte ratio (MLR) [29]. Furthermore, the generalisability of the model is a concern as it was developed using data from patients undergoing spinal fusion surgery in a tertiary level A hospital in Liuzhou City, Guangxi Zhuang Autonomous Region, China. It is possible that the results may not be applicable to other populations, such as hospitals in different countries or regional healthcare systems with different levels of hierarchy. Although internal and external validation through bootstrapping confirmed the robustness of the model, external validation was only performed on hospitals at the same level in the same region. Therefore, external validation in different clinical settings and populations is still needed to fully establish its generalisability. It is recommended that future studies focus on validating the model in multiple centres with different patient populations. In addition, it should be noted that the data were collected one year ago, and changes in clinical practice, patient management strategies or healthcare policies may affect model performance over time. Ongoing monitoring and periodic recalibration are therefore required to maintain accuracy.

Conclusions

Certain factors, such as blood transfusion, type of surgery, use of TXA, diabetes, electrolyte disturbance, BMI, surgical procedure performed, the number of preoperative diagnoses and operation time, increase the risk of prolonged LOS. These risk factors were used as basis to develop an easy-to-apply prediction model to detect spinal fusion patients who are at risk of prolonged LOS. Early risk stratification of patients is a beneficial outcome, and future large-scale multi-centre research are necessary to evaluate the efficacy of this model.