1. Introduction
Intertrochanteric fracture is considered one of the most concerning health conditions in the elderly population. Over the years, the incidence of this fracture type has increased concordantly with aging [
1]. Patients with intertrochanteric fractures can be treated with either an operative or non-operative approach. However, since operative treatment was proven to be significantly superior in reducing morbidity and mortality compared to non-operative treatment, it is currently recommended as the standard treatment for intertrochanteric fracture [
2]. Unfortunately, approximately half of the patients who underwent surgical operation with an uneventful fracture healing still had a poor ambulatory status [
2].
As not all patients would achieve good ambulatory status after surgery, prediction of postoperative functional outcomes is clinically important and might be helpful for both physicians and patients in finding common ground on the postoperative care plan (e.g., deciding whether an intensive postoperative rehabilitation program should be prescribed, or an early discharge plan should be considered instead) [
3,
4]. Several preoperative factors were identified to predict treatment outcomes in patients with intertrochanteric fractures, such as patients’ baseline condition and pre-injury ambulatory function [
5,
6]. Some intraoperative parameters, such as reduction alignment and implant position, have previously been shown to be significant in predicting fixation failure [
7,
8].
Although these prognostic factors were associated with postoperative treatment outcomes, they were usually considered separately. In other words, they could not provide individual predictions, which would be more practical and relevant in clinical practice. One solution is to create clinical prediction rules (CPR) using statistical modeling, which would utilize the information from several predictors for estimating the probability, or the likelihood, of postoperative outcomes specifically for each patient [
4]. For patients with intertrochanteric fractures, individualized predictions might be beneficial in at least two circumstances. First, primary care physicians or emergency physicians, who were usually the first to encounter the patients, could provide a rough estimation of postoperative outcomes to the patients and their families using only preoperative information [
9]. Second, orthopaedic surgeons could use both preoperative and intraoperative information to predict individual treatment outcomes and use these predictions as guidance during fracture reduction and fixation adjustment to achieve satisfactory treatment outcomes [
10].
Currently, only a limited number of CPR were available to predict functional outcomes in patients with intertrochanteric fractures [
11,
12]. In addition, most of the tools were developed to predict the ability to withstand mechanical failure after fracture fixation, or fixation failure [
10,
13,
14]. Even though fixation failure is an important endpoint that should not be overlooked, predicting postoperative functional outcomes is also clinically relevant and might be more straightforward to the patients [
15,
16]. Thus, we intended to develop CPR using preoperative and intraoperative information to predict patients’ ability to ambulate one year after surgery, a clinically relevant functional outcome.
3. Results
During the study period, 266 patients diagnosed with intertrochanteric fracture were included. Of those, we excluded 41 patients who have passed away, three patients who suffered from high energy injury mechanism, and one patient with pathological fracture. Finally, 221 patients were included, with 160 (72.4%) patients with an NMS ≥ 5. The comparison of baseline clinical characteristics, fracture configuration, and surgical-related parameters between patients with good and poor ambulatory status at one year is presented in
Table 1.
For the preoperative model, sex, BMI, CCI, and pre-injury NMS were included in the multivariable logistic regression. After applying the MFP algorithm, only the pre-injury NMS parameter was fitted and transformed into the second-degree fractional polynomial (FP2) terms. In contrast, the rest were fitted with the linear term (
Appendix A:
Figure A1).
Table 2 presents the regression coefficient, 95% confidence intervals, and
p-value of the transformed covariates. An AuROC of the derived model revealed an acceptable discriminative ability of 0.77 (95% CI 0.70 to 0.85) (
Figure 2).
The intraoperative model was derived using the predicted probability from the preoperative model in combination with the preselected surgical-related parameters, including the AO/OTA classification, the lateral wall thickness, the NSA, the displacement, the CalTAD, Parker’s ratio, and the type of fixation implant. All predictors were fitted with linear terms without covariate transformation. The coefficients of each parameter with 95% CI and
p-value are shown in
Table 3. After incorporating surgical-related parameters into the preoperative model, the discriminative ability significantly increased by approximately 6% to an AuROC of 0.83 (95% CI 0.77 to 0.88) (
p = 0.021) (
Figure 2). This improvement represents the added discriminative value of surgical-related parameters to preoperative predictors.
The calibration of both models was presented with calibration plots (
Figure 3). Pearson’s goodness-of-fit statistics were insignificant for preoperative and intraoperative models (
p = 0.809 and
p = 0.693, respectively). Internal validation with bootstrap re-sampling revealed an optimism of 0.01 (range −0.07 to 0.11) and 0.04 (range −0.04 to 0.14) with an estimated shrinkage factor of 0.92 (95% CI 0.90 to 0.94) and 0.80 (95% CI 0.78 to 0.82) for the preoperative and intraoperative model, respectively.
4. Discussion
In the present study, we developed the CPR for both preoperative and intraoperative prediction of postoperative functional outcome at one year for patients with intertrochanteric fractures. The preoperative model consists of four predictors: sex, BMI, CCI, and pre-injury NMS. For the intraoperative model, surgical-related parameters, including fracture configuration, fracture reduction parameters, fixation quality, and type of implant, were added to the preoperative model. It was found that these intraoperative factors significantly improved the discriminative ability of the preoperative model in predicting postoperative ambulatory status at one year.
Previously, few CPR were developed to predict postoperative outcomes in patients with intertrochanteric fractures. Tanaka et al. developed a CPR for predicting declination of activity of daily living (ADL) at six months after operation [
11]. The CPR was based entirely on patients’ clinical characteristics and did not consider the relevant surgical parameters. Subsequently, another CPR was developed by Murena et al. to predict implant failure using surgical-related parameters [
10]. In this study, the IT-AP tool was developed to predict postoperative functional outcomes in terms of ambulatory status using both clinical and surgical-related parameters.
The IT-AP tool is composed of two models: the preoperative and intraoperative models. The preoperative model was proven to deliver an acceptably accurate prediction of the patient’s ambulatory status after surgery, which could be used by attending physicians for risk communication in primary or emergency care settings [
33]. Within the preoperative model, the patient’s pre-injury ambulatory status was the strongest predictor of having a good postoperative functional outcome. Pre-injury ambulation ability was an independent predictor of functional outcome [
7] and was also included in a previous CPR. In addition, as all predictors included within the preoperative model were chosen based on clinical availability, the model was highly pragmatic.
The intraoperative model was developed by incorporating surgical-related parameters with the predicted probabilities from the preoperative model. The model provides the likelihood of achieving a good one-year postoperative functional outcome, which can be used intraoperatively by orthopaedic surgeons in determining the current quality of reduction and fixation. All of the included parameters were preselected according to the previous literature [
8,
13,
14]. An anatomical to a slightly valgus reduction of the fracture was associated with a high rate of fracture union [
13]. The positive cortical reduction improves the fixation stability by limiting the fracture displacement [
13,
34].
Although each of the surgical-related parameters did not show statistically significant association with postoperative functional outcome, we included these parameters within our intraoperative model as it was previously proven to be effective in predicting fracture stability, which is one of the precursors of good postoperative functional status [
25]. As a result, when surgical-related parameters were included with the preoperative predictions, there was a significant improvement in the discriminative ability of the intraoperative model. Our findings support the importance of considering both the patients’ initial clinical information and the intraoperative surgical parameters in predicting postoperative functional outcomes.
To demonstrate how the IT-AP tool would fit in clinical practice, we present a case of a 75-year-old male who presented with left hip pain and was subsequently diagnosed with an intertrochanteric fracture. To predict the likelihood of having good ambulatory status one year after the operation, the attending physician filled in all required predictor information into the IT-AP tool as shown in
Figure 4 (
Appendix C:
Figure A4,
Table A3). Then, the IT-AP tool under the preoperative model option estimated the LHR of having good ambulatory status after receiving operation at 4.7 (
Appendix C:
Figure A5a), which means that it was more likely than not that the patient would be able to ambulate at one year. In this case, the patient was scheduled for closed reduction and internal fixation with an intramedullary device (
Appendix C:
Figure A4a,b). After an initial adjustment, the orthopaedic surgeon wanted to know if the “current” adjustment was adequate. Surgical-related parameters were entered into the IT-AP tool under the intraoperative model option to estimate the LHR of ambulation considering fixation characteristics. The LHR was estimated at 3.70, representing a suboptimal surgical treatment (
Appendix C:
Figure A5b). The surgeon was then informed by the IT-AP tool that the “current” fixation and reduction characteristics were still inadequate and that further adjustment was still required (
Appendix C:
Figure A4c,d). Finally, after the surgeon performed additional surgical adjustments, the IT-AP tool revealed a LHR of 10.59, which was substantially higher than the predicted LHR from the preoperative model. Therefore, the surgeon decided to proceed with the skin closure and finish the operation (
Appendix C:
Figure A5c).
There are several strengths to this study. Our study was one of few studies to develop CPR for predicting the postoperative functional outcome [
11]. The MFP algorithm preserves the nature of continuous predictors, which prevents significant loss of information [
31]. Moreover, the simplified presentation of our CPR allows clinicians and specialists to calculate the individualized prediction for each patient easily. However, there are some limitations to be addressed. First, our CPR was developed using a relatively small cohort of patients, which may cause model overfitting. Nevertheless, only a minimal amount of model optimism was observed. Second, this study was susceptible to selection bias for excluding patients who passed away. For that reason, the study cohort might not be able to reflect the entire intended population, and therefore predicted odds, and a positive likelihood ratio, was reported instead of the predicted probability to avoid the misinterpretation of the predicted probability. However, it is important to note that the predicted odds, or the predicted LHR, from our model could overestimate the likelihood of ambulation. The degree of bias is unknown and needs to be proven in the validation study. Third, telephone interviews to collect outcome data may give rise to recall and interviewer biases. However, these biases were minimized by using a structured interview with a blinded interviewer [
29]. Lastly, the apparent model performance reported might not generalize to other settings. Thus, external validation should be conducted before being implemented in practice.
While it is true that the surgery would be performed to the best of the surgeon’s ability, the definition of “the best” is not well described. In most cases, an acceptable range of an adequate quality of fixation and reduction varied among different surgeons [
35,
36]. From our perspective, the quality of surgery should be determined differently across each individual. An individualized prediction could help determine the most appropriate surgical-related parameters for each patient in order to achieve good postoperative ambulatory status. Therefore, we believe that our tool would benefit clinical practice by tailoring treatment for each patient.