1 Introduction

Currently, the majority of outcomes following total hip arthroplasty (THA) are excellent and mortality is low, at less than 1% within both 30 and 90 days of procedure.^1–5 However, due to the sheer volume of THA cases, the total number of deaths following THA is significant, and its risk factors should thus be explored. To our knowledge, few studies have examined the specific risk factors that increase mortality following primary THA. In one study, Ibrahim et al. demonstrate that a patient's race and ethnicity do not affect 30-day mortality rate following primary THA.² In another study, Hunt et al. argue that a posterior surgical approach, mechanical thromboprophylaxis, chemical thromboprophylaxis with heparin, and spinal anesthesia are independently associated with an increased 90-day mortality following primary THA.⁶

Artificial Neural Networks (ANN) are computational models that have recently been used in medicine to model and evaluate risk factors for various patient complications.^7,8 Based off of biological neural networks, studies have shown that ANNs can more accurately analyze health risk factors than currently employed statistical methods.^9–11 For example, an ANN examining 17 risk factors predicted inpatient mortality in the intensive care unit with 94% accuracy.⁷ In another study that examined complications in posterior lumbar spine fusion, ANNs predicted cardiac complications, wound complications, venous thromboembolism, and mortality better than alternative methods.⁸ Thus, the purpose of this study is 1) to develop and validate an ANN model for primary THA, and 2) to identify the most important risk factors leading to increased 30-day patient mortality following primary THA.

2 Methods

Data was collected from the American College of Surgeons National Surgery Quality Improvement Program (ACS-NSQIP). ACS-NSQIP is a large multicenter surgical database that is prospectively collected by trained surgical clinical reviewers at over 500 participating U.S., Canadian, United Kingdom, Republic of Korea, Pakistan, Lebanon Australia, Saudi Arabia, Jordan, Germany, Singapore, Taiwan, Philippines, Guam, Cuba, Italy, Spain and United Arab Emirates institutions. Rigorous collection of data and auditing of the data by the ACS has led to considerably high inter-rater reliability.^12,13

Subjects for this study were identified using Current Procedural Terminology codes. Primary Current Procedural Terminology codes 27130 was used to identify patients receiving primary total hip arthroplasties for osteoarthritis. Subjects were excluded for: any missing data, additional procedures, ASA classification V, septic, and emergent cases. After applying the exclusion criteria, 77,145 patients remained for analysis. The cohort was further stratified by presence of mortality within 30 days postoperatively.

Statistical Package for the Social Sciences (SPSS; Version 26; Armonk, NY) software was used for all statistical analyses. Chi-square and independent-t test analyses were performed to compare the two groups. A p-value<0.05 is considered statistically significant. Statistically significant variables were used for the ANN model. The settings for the ANN model were the following:

Initial mode choice- Multilayer Perception.

Variable window- Nominal variables entered as “Factors,” Ordinal variables entered as “Factors,” Continuous variables entered as “Covariates,” and outcome target variable entered as “Dependent variable”

Partitioning of the cases- All training of the model and cross validation were carried out as 70% of the cases assigned to training and 30% of the cases assigned to testing.

Rescaling covariate- Standardized option.

Architecture- Minimum number nodes in hidden layer: 1; Maximum number of nodes in hidden layer: 50.

Training criteria- Type of training: Batch; Optimization algorithm: scaled conjugate gradient; Initial Lambda: 0.0000005; Initial Sigma: 00005; Interval center: 0; Interval offset: 0.5.

User missing values- Exclude.

Output- Network structure: Description, diagram; Network performance: Model summary, Classification results, ROC curve; Case processing summary; Independent variable importance analysis.

Stopping rules- Maximum steps without a decrease in error: 5.

Other choices were default options.

3 Results

3.1

3.1 Patient cohort

In total, 77,145 patients were included in this analysis. Among those, 114 patients (0.15%) patients died within 30 days. Patients who died within 30 days were more likely to be older (74.9 years old vs 65.2 years old, p < 0.001), and have a lower BMI (28.8 vs 30.3, p = 0.009; Table 1). In addition, patients that died within 30 days were more likely to have diabetes (p = 0.002), have dyspnea (p < 0.001), are less independent (p < 0.001), have COPD (p = 0.017), have congestive heart failure (p < 0.001), have hypertension (p < 0.001), need dialysis (p < 0.001), have a bleeding disorder (p < 0.001), and have a preoperative transfusion within 72 h prior to surgery (p < 0.001; Table 1). Lastly, patients that died within 30 days were more likely to have a higher ASA classification (p < 0.001), receive general anesthesia (p = 0.015), have a lower preoperative hematocrit (p < 0.001), higher INR (p < 0.001), and have a longer operative time (p < 0.001; Table 1). The remaining demographic differences are further illustrated in Table 1.

Table 1 Demographics and clinical characteristics among patients undergoing total hip arthroplasty.

Patient characteristics	Total		Mortality		No Mortality		P-value
	77145		114		77031
	N	%	N	%	N	%
Sex							0.631
Female	42439	55.0%	60	52.6%	42269	54.9%
Male	34816	45.1%	54	47.4%	34762	45.1%
Race							0.751
White	65918	85.4%	99	86.8%	65819	85.4%
Black or African American	6678	8.7%	10	8.8%	6668	8.7%
Hispanic	2737	3.5%	2	1.8%	2735	3.6%
Asian	1236	1.6%	3	2.6%	1233	1.6%
Native American or Alaskan	320	0.4%	0	0.0%	320	0.4%
Hawaiian or Pacific Islander	256	0.3%	0	0.0%	256	0.3%
Age (years old)	65.3 ± 11.0		74.9 ± 11.0		65.2 ± 11.0		<0.001
BMI	30.3 ± 6.4		28.8 ± 7.5		30.3 ± 6.4		0.009
Anesthesia							0.015
General	43757	56.7%	81	71.1%	43676	56.7%
Neuraxial	24846	32.2%	22	19.3%	24824	32.2%
Regional	1208	1.6%	1	0.9%	1207	1.6%
MAC/IV	7334	9.5%	10	8.8%	7324	9.5%
Diabetes							0.002
No Diabetes Mellitus	67643	87.7%	98	86.0%	67545	87.7%
Non-insulin dependent Diabetes Mellitus	7343	9.5%	7	6.1%	7336	9.5%
Insulin-dependent Diabetes Mellitus	2159	2.8%	9	7.9%	2150	2.8%
Smoke	114	0.1%	8	7.0%	106	0.1%	0.077
Dyspnea							<0.001
No dyspnea	73627	95.4%	99	86.8%	73528	95.5%
Dyspnea on exertion	3343	4.3%	15	13.2%	3328	4.3%
Dyspnea at rest	175	0.2%	0	0.0%	175	0.2%
Functional status							<0.001
Independent	75577	98.0%	97	85.1%	75480	98.0%
Partially dependent	1519	2.0%	13	11.4%	1506	2.0%
Dependent	49	0.1%	4	3.5%	45	0.1%
Chronic Obstructive Pulmonary Disorder	3046	3.9%	12	10.5%	3034	3.9%	<0.001
Congestive Heart Failure	237	0.3%	3	2.6%	234	0.3%	<0.001
Hypertension needing medication	45006	58.3%	87	76.3%	44919	58.3%	<0.001
Renal Failure	23	0.0%	0	0.0%	23	0.0%	1.000
Maintenance Dialysis or dialysis within 14 days	150	0.2%	4	3.5%	146	0.2%	<0.001
Chronic steroid use	2550	3.3%	6	5.3%	2544	3.3%	0.283
Weight loss more than 10% within 6 months	168	0.2%	0	0.0%	168	0.2%	1.000
Bleeding disorder	2057	2.7%	16	14.0%	2041	2.6%	<0.001
Preoperative Transfusion <72 h	89	0.1%	4	3.5%	85	0.1%	<0.001
ASA classification							<0.001
1 or 2	43505	56.4%	24	21.1%	43481	56.4%
3 or 4	33640	43.6%	90	78.9%	33550	43.6%
Preoperative Hematocrit	40.96 ± 4.18		38.45 ± 5.55		40.96 ± 4.18		<0.001
Preoperative INR	1.03 ± 0.26		1.11 ± 0.31		1.03 ± 0.26		<0.001
Operative Time (min)	93.65 ± 38.6		95.6 ± 39.8		93.64 ± 38.6		<0.001

The current study's ANN model utilized the following variables when constructing the neural network: anesthetic type, diabetes, dyspnea status, functional status, COPD, congestive heart failure, hypertension, dialysis, bleeding disorder, ASA classification, age, BMI, preoperative hematocrit, preoperative INR, and operative time. The model had an incorrect prediction rate of 0.2%. The AUC of the current model was 80.0% (Fig. 1). The top five most important independent variables the model used to predict mortality were preoperative INR (importance = 0.162), age (importance = 0.152), BMI (importance = 0.147), operative time (importance = 0.143), and preoperative hematocrit (importance = 0.094; Table 2, Fig. 2).

Close

Fig. 1 ROC curve of the Artificial Neural Network model in predicting 30-day postoperative mortality after primary total hip arthroplasty; DOpertoD- Mortality.

Export to PPT

Table 2 Importance chart of the Artificial Neural Network model.

Independent Variable Importance	Importance	Normalized Importance
Anesthetic type	0.033	20.50%
Diabetes	0.026	16.00%
Dyspnea	0.027	16.50%
Functional Status	0.021	13.10%
Chronic Obstructive Pulmonary Disorder	0.017	10.40%
Congestive Heart Failure	0.015	9.00%
Hypertension requiring medication	0.014	8.80%
Maintenance dialysis or dialysis less than 14 days	0.064	39.50%
Bleeding disorder	0.037	22.80%
Preoperative transfusion less than 72 h	0.01	6.50%
ASA Classification	0.038	23.20%
Age	0.152	93.70%
Preoperative hematocrit	0.094	57.70%
Preoperative INR	0.162	100.00%
BMI	0.147	90.90%
Operative Time	0.143	87.90%

Close

Fig. 2 Importance of factors used by the Artificial Neural Network model to determine mortality; PRINR-preoperative INR; AGE-age; BMI- Body Mass Index; OPTIME-operative time; PRHCT-preoperative hematocrit; DIALYSIS- maintenance dialysis or dialysis within 14 days; ASACLAS_USE- American Society of Anesthesiologists Class; BLEEDIS- bleeding disorder; ANESTHES_USE-anesthetic type; DYSPNEA-dyspnea; FNSTATUS2- functional status; HXCOPD-chronic obstructive pulmonary disorder; HXCHF- congestive heart failure; HYPERMED-hypertension requiring medication; TRANSFUS- preoperative transfusion <72 h.

Export to PPT

4 Discussion

This study utilized ANN and computational modeling to show that the top five most important independent variables that predicted increased mortality following primary THA were higher preoperative INR, greater age, lower BMI, increased operative time, and lower preoperative hematocrit. Although ANN has been used in to identify adverse events, increased costs, and longer length of stay after total hip and knee arthroplasty, to our knowledge, this is the first study to identify risk factors for mortality after primary THA using this modeling method.^8,9,14,15

Regarding perioperative anticoagulation following total joint arthroplasty, the risk of thromboembolism should be weighed against the risk of bleeding. In their study, Hunt et al. determined that both mechanical and chemical thromboprophylaxis with heparin was associated with a statistically significant reduction in mortality following primary THA.⁶ In our study, we demonstrate that a higher preoperative INR is the leading predictor of increased mortality following THA. Similarly, in their study of 1047 primary THA's, McDougall et al. demonstrate that patients on warfarin have a higher rate of complications including deep joint infection, hematoma, and superficial infections.¹⁶ Importantly, the increased INR analyzed in our study may be reflective of other factors such as patient comorbidities, medications that lead to increased INR (including anticoagulants as well as non-anticoagulant medications that can increase INR), or disorders that necessitate long-term anticoagulation. This highlights the importance of preoperative medical optimization for patients undergoing elective joint arthroplasty and further emphasizes the importance of carefully evaluating preoperative INR before surgical clearance.

In addition, our study demonstrated that age is an important predictor of 30-day mortality following THA, as we found that, on average, patients who died following surgery were 10 years older than patients who survived. Similarly, Hunt et al. demonstrate an association between increased age and 90-day mortality following primary THA.⁶ One possible explanation is due to the increased prevalence of chronic conditions with age, as over 80% of Americans over the age of 65 have a chronic condition.¹⁷ We found that the presence of several chronic conditions, notably diabetes, dyspnea, COPD, CHF, HTN, and dialysis, were all independently associated with increased 30-day mortality following THA.

We found that patients with a lower BMI had a statistically significant increased 30-day mortality following THA (p = 0.009). This finding was supported by Hunt et al., who also noted that being overweight was associated with a statistically significant decrease in 90-day mortality following THA.⁶ In their study, they noted that this finding could be explained as the majority of immediate deaths following THA are due to cardiovascular causes, and that in patients with cardiovascular disease, those with a greater BMI have a decreased mortality.^1,6 The association of lower BMI and increased mortality has also been demonstrated in the hip fracture literature.¹⁸ Termed the “obesity paradox,” studies have reported a correlation with higher BMI and increased survival after hip fracture in the geriatric population, likely due to the negative effects of underweight and malnutrition seen in patients with lower BMI.¹⁸ This “obesity paradox,” is in agreement with the results of the present study as well, as lower BMI was associated with increased 30-day mortality after THA. These findings may highlight the importance of nutritional optimization in patients who are malnourished or underweight prior to undergoing primary THA.

Lastly, we found that an increased operative time and lower preoperative hematocrit were associated with increased 30-day mortality following THA. Although Surace et al. demonstrate that longer operative times have a significant association with short-term complications following THA, they fail to report a statistically significant association with operative time and 30-day mortality following THA.¹⁹ Importantly, the results of the present study do not determine any causal relationship between longer operative time and increased mortality risk. However, there may be several explanations for this association of longer operative time and mortality, such as increased case complexity, lower surgeon volume, or greater intraoperative bleeding due to medical comorbidities or other patient factors. In addition, these findings are supported by other recent studies which have demonstrated that patients with anemia are at increased risk of postoperative complications and mortality following total hip arthroplasty.^20,21

Our study thus validates the use of an Artificial Neural Network to identify risk factors that are important in predicting mortality following THA. Previous studies have demonstrated that ANN's have efficacy in predicting mortality in intensive care unit patients.⁷ In orthopedic surgery, ANN's have been used to predict outcomes following posterior lumbar spine fusions and predicting length of stay and costs following total knee and hip arthroplasty.^8,14,15 A neural network model has also been shown to be capable of identifying and classifying knee osteoarthritis as accurately as a fellowship-trained arthroplasty surgeon.²² In addition, Magneli et al. showed that a machine learning model relying on administrative data was more accurate at predicting 90-day adverse events after hip arthroplasty than a model based on international classification of diseases (ICD) codes.⁹ To our knowledge, this is the first study that demonstrates the utility of an ANN in identifying specific patient factors that contribute to increased mortality following primary THA. As demonstrated above, we use the ANN to not only validate several of the previously known risk factors for increased mortality following THA, but also identify several risk factors that have not yet been reported in the literature. Notably, several of these factors, such as hematocrit and INR, can be evaluated preoperatively through laboratory tests and optimized through medical management. We argue that these factors should be carefully considered before patients are medically cleared for primary THA.

There are several notable limitations to our study. First, the NSQIP database is only able to assess complications up to one month postoperatively. Thus, we are unable to report any mortality events after this time period, leading to an underreporting of 90-day or long-term mortality following THA. Another limitation of the study arises from potential inaccurate coding of patient data and mortality. Despite this limitation, prior studies have demonstrated a high inter-rater reliability with regards to data collection when using the NSQIP database.^12,13 Lastly, the NSQIP only provides data from participating institutions, and leaves the possibility of excluding a large number of patients undergoing primary THA. Thus, our study may not be generalizable to all patients undergoing primary THA; however, our sample size of 77,145 patients lends significant power to our study. Moreover, we found that our study population demonstrated an overall 30-day mortality rate of 0.15% (114 patients), a finding similar to that reported in the literature.¹ One strength of the NSQIP is that it provides a large sample size which is needed to analyze rare events, such as mortality after THA, that cannot be studied as well using single surgeon, or even multicenter cohorts.

5 Conclusion

Overall, with the rapid increase in amount of THA being performed in the United States, it will be key to limit all-cause mortality following these procedures. Our study has developed and utilized an Artificial Neural Network (ANN) to identify five factors that predict 30-day mortality following THA: a higher preoperative INR, greater age, lower BMI, increased operative time, and lower preoperative hematocrit. We argue that these factors should be acknowledged and evaluated before patients undergo THA. If these precautions are taken, our study demonstrates that the mortality rate following primary THA could decrease even further.

Funding

No funds, grants, or other support was received.

Conflicts of interest/competing interests

The authors have no conflicts of interest to declare that are relevant to the content of this article.

Availability of data and material

Not applicable.

Code availability

Not applicable.

Financial interests

The authors declare they have no financial interests.

Ethics approval

Institutional review board approval was not required for this study as all data is de-identified and public.