Share this article on:

Applying Latent Class Analysis to Risk Stratification for Perioperative Mortality in Patients Undergoing Intraabdominal General Surgery

Kim, Minjae MD, MS; Wall, Melanie M. PhD; Li, Guohua MD, DrPH

doi: 10.1213/ANE.0000000000001279
Healthcare Economics, Policy, and Organization: Research Report

BACKGROUND: Perioperative risk stratification is often performed using individual risk factors without consideration of the syndemic of these risk factors. We used latent class analysis (LCA) to identify the classes of comorbidities and risk factors associated with perioperative mortality in patients presenting for intraabdominal general surgery.

METHODS: The 2005 to 2010 American College of Surgeons National Surgical Quality Improvement Program was used to obtain a cohort of patients undergoing intraabdominal general surgery. Risk factors and comorbidities were entered into LCA models to identify the latent classes, and individuals were assigned to a class based on the highest posterior probability of class membership. Relative risk regression was used to determine the associations between the latent classes and 30-day mortality, with adjustments for procedure.

RESULTS: A 9-class model was fit using LCA on 466,177 observations. After combining classes with similar adjusted mortality risks, 5 risk classes were obtained. Compared with the class with average mortality risk (class 4), the risk ratios (95% confidence interval) ranged from 0.020 (0.014–0.027) in the lowest risk class (class 1) to 6.75 (6.46–7.02) in the highest risk class. After adjusting for procedure and ASA physical status, the latent classes remained significantly associated with 30-day mortality. The addition of the risk class variable to a model containing ASA physical status and surgical procedure demonstrated a significant increase in the area under the receiver operator characteristic curve (0.892 vs 0.915; P < 0.0001).

CONCLUSIONS: Latent classes of risk factors and comorbidities in patients undergoing intraabdominal surgery are predictive of 30-day mortality independent of the ASA physical status and improve risk prediction with the ASA physical status.

Supplemental Digital Content is available in the text.Published ahead of print April 22, 2016

From the *Department of Anesthesiology, Columbia University Medical Center, New York, New York; Department of Biostatistics, Mailman School of Public Health, Columbia University, New York, New York; and Department of Epidemiology, Mailman School of Public Health, Columbia University, New York, New York.

Accepted for publication February 8, 2016.

Published ahead of print April 22, 2016

Funding: Supported by a grant from the International Anesthesia Research Society.

The authors declare no conflicts of interest.

This work was presented as an abstract at the International Anesthesia Research Society Annual Meeting, March 23, 2015.

Supplemental digital content is available for this article. Direct URL citations appear in the printed text and are provided in the HTML and PDF versions of this article on the journal’s website.

Reprints will not be available from the authors.

Address correspondence to Minjae Kim, MD, MS, Department of Anesthesiology, Columbia University Medical Center, 622 West 168th St., PH 5, Suite 505C, New York, NY 10032. Address e-mail to mk2767@cumc.columbia.edu.

Perioperative risk stratification is a critical aspect in the management of patients presenting for surgical procedures. There are many methods used to assess perioperative mortality risk, each with specific benefits and limitations. The most widely used perioperative risk stratification measure is the ASA physical status,a,1–3 which was originally developed to standardize data collection for statistical analysis4 and not as a measure of operative risk, because this was subject to variability depending on factors such as the surgical procedure, skill of the practitioners, and anesthetic management.

Other perioperative stratification tools include comorbidity indexes, such as the Charlson Comorbidity Index5,6 and the Elixhauser index.7 These indexes include individual risk factors and generally provide an indication of the relative number of conditions that are present, but they do not specifically assess for interactions between comorbidities that may have detrimental effects on clinical outcomes.8 To gain a more thorough understanding of the effects of comorbidities, it is necessary to view comorbid status as the totality of all of the conditions that are present and not simply as the additive effects of single, independent factors.

In the current study, we use a novel approach to assess the risk of perioperative mortality based on the entirety of a patient’s comorbidities and risk factors using latent class analysis (LCA), a statistical modeling technique used to identify the groups of subjects that are similar with respect to a set of observed characteristics.b LCA involves the use of latent variables, which are unobserved variables that cannot be directly measured but instead are estimated from other variables that can be measured.9 In this study, using LCA, we incorporated comorbidity and risk factor data of patients presenting for general surgery to model the comorbid state as a latent variable. We hypothesized that this approach will identify patterns of comorbidities and risk factors that are present in the population and that also contribute to increased perioperative risk, allowing for a better understanding of the relationships between these comorbidities and perioperative mortality. Although LCA is not a new technique, its utilization in the context of perioperative risk assessment by comorbid status has not been described. We compared perioperative mortality risk assessment using LCA methods with the ASA physical status. We analyzed data from the American College of Surgeons National Surgical Quality Improvement Program (ACS-NSQIP), a large, multicenter database of surgical outcomes from hospitals throughout North America.

Back to Top | Article Outline

METHODS

Background on Latent Variables

Latent variable models include unobserved random variables that are modeled from observed variables, such as factor, item response, latent class, and structural equation models.10 A latent variable is one that cannot be directly measured, such as happiness11 and intelligence,12 but is instead inferred from other variables that can be measured.13 LCA is a statistical modeling technique used to identify the groups of subjects that are similar with respect to a set of observed characteristics.b LCA is widely used in behavioral and social science fields such as psychiatry where disease states cannot be directly measured, such as in identifying different subtypes of patients with schizophrenia14 or posttraumatic stress disorder.15 Applications of LCA in medical research, including determining the probability of disease based on several observed factors when the disease itself does not have standard diagnostic measures,16,17 have been described.18 Further information on LCA and latent variable methods are available,c,d including software packages used for latent variable analysis.e

We cannot directly measure a patient’s comorbid state, but it can be modeled as a latent variable using data on preoperative comorbidities and risk factors. Using a sample of patients undergoing intraabdominal general surgery, we used LCA to determine whether there are distinct subtypes of patients with similarities in their comorbidities and risk factors, that is, to group patients into different categories based on the severity of their comorbidities and risk factors.

Back to Top | Article Outline

Data

This study was not subject to review by the Columbia University Medical Center IRB (New York, NY) because it did not require access to protected health information. The ACS-NSQIPf is a validated, prospectively collected national data set aimed at improving surgical quality and outcomes.19 The data collected included demographic characteristics, presurgical comorbidities, intraoperative variables, and 30-day postoperative morbidity and mortality. All data were carefully reviewed by each site’s surgical clinical reviewer and centers not meeting specific criteria for quality were removed from the data set. The systematic sampling process and criteria for maintaining the high quality of the data set have been described previously.20

Back to Top | Article Outline

Patient Selection

We selected patients undergoing intraabdominal general surgery using data from the 2005 to 2010 ACS-NSQIP participant use files. The Clinical Classifications Software for Services and Procedures (Agency for Healthcare Research and Quality, Rockville, MD)g was used to classify procedures based on the primary Current Procedural Terminology (American Medical Association, Chicago, IL) code. Fifteen intraabdominal general surgery categories were identified for consideration (Appendix 1), resulting in an initial sample of 630,542 records. Most categories refer to a specific type of procedure, but of the body system categories, “other operating room gastrointestinal procedures” consists mainly of hepatobiliary and pancreatic procedures, whereas “other operating room lower gastrointestinal therapeutic procedures” consists of intestinal and colorectal cases that do not fit a specific procedure category. We have previously demonstrated that regression models using the Clinical Classifications Software for Services and Procedures to categorize procedures were robust and did not introduce additional variation compared with using individual Current Procedural Terminology codes.21 Using similar methodology, we confirmed this to be case for the current analysis (data not shown). Patients classified as outpatient were excluded (N = 164,365) because they have a low risk of adverse perioperative outcomes. No further exclusion criteria were used, and the final analysis cohort had 466,177 observations.

We created an additional cohort using data from the 2011 ACS-NSQIP participant use file to determine whether LCA methodology could be applied to an additional sample to reproduce the latent risk classes. After applying the same criteria described above, the cohort consisted of 61,593 records.

Back to Top | Article Outline

Baseline Demographic and Operative Variables

Data on risk factors and comorbidities were collected directly from the ACS-NSQIP data set. Detailed descriptions for specific variables are available.h Age was categorized as: ≤40, 40 to 50, 50 to 60, 60 to 70, 70 to 80, and >80 years. Body mass index (BMI) was calculated from height and weight data and categorized based on World Health Organization guidelines22: <18.5, 18.5 to 25, 25 to 30, 30 to 35, and >35 kg/m2. The estimated glomerular filtration rate (mL/min/1.73 m2) was calculated using the Chronic Kidney Disease Epidemiology Collaboration formula incorporating creatinine, sex, age, and race23 and categorized into groups corresponding to the stages of chronic kidney disease (CKD)24: <15, 15 to 30, 30 to 60, 60 to 90, >90, or missing. The categorization for hematocrit (%) was determined empirically by plotting the decile of hematocrit against the log risk of mortality to visualize the relationship between the variables; hematocrit was categorized as <34, 34 to 38, >38, or missing. Not all patients require a full laboratory workup before surgery and a missing value may be an important prognostic indicator for perioperative morbidity and mortality. A history of cancer was determined as described previously.21 Data on other comorbidities and patient characteristics were collected directly from the data set.

Back to Top | Article Outline

LCA Model

Figure 1

Figure 1

LCA models were used to categorize patients into distinct risk classes based on 31 risk factors and comorbidities known before surgery (Fig. 1). For each LCA model, the model parameters are estimated based on the specified number of classes. Once the estimated parameters are known, the probability of belonging to each class (or posterior probability) can be calculated for each patient. For instance, a 3-class model might return posterior probabilities of 90%, 8%, and 2% for a particular patient. Models with increasing numbers of latent classes were fit, and determination of the best number of classes was based on comparison of Bayesian Information Criterion (BIC)25 and clinical interpretation.26 After determining the number of latent classes, each patient was assigned to a latent class based on the highest posterior probability. To confirm the validity of this approach, we created 1000 bootstrapped samples where the assigned latent risk class of an individual patient was allowed to vary based on the posterior probabilities. Data were missing for <1% of all variables except for BMI, where 3.7% of records had missing data. Full-information maximum likelihood was used to account for missing data in the LCA model.27 For the 2011 ACS-NSQIP cohort, parameter estimates from the original LCA model were used to estimate the probabilities of class membership. LCA was performed using Mplus software version 7 (Muthén & Muthén, Los Angeles, CA).

Back to Top | Article Outline

ASA Physical Status

The ASA physical status was determined from the data set, with classifications ranging from ASA physical status I to ASA physical status V (Appendix 2).

Back to Top | Article Outline

Clinical End Points

The follow-up period for patients participating in the ACS-NSQIP is 30 days after their procedure, and the primary outcome was 30-day mortality.

Back to Top | Article Outline

Statistical Analysis

Poisson regression modeling with robust variance28 was performed to assess the relative risk of mortality among the different clinical outcomes groups using risk ratios (RRs). Calibration for each model was assessed by dividing the sample into deciles of mortality risk and plotting predicted mortality versus observed mortality. Model discrimination was assessed using the area under the curve (AUC) of the receiver operator characteristic curve. Statistical analyses were performed using SAS Software version 9.4 (SAS Institute, Cary, NC). In all analyses, statistical significance was determined with a P value of <0.05.

Back to Top | Article Outline

RESULTS

Determination of the Latent Risk Classes

We fit latent class models based on 31 comorbidities and risk factors (Fig. 1), ranging from 2 classes to 10 classes. With each increment in the number of latent classes, the fit criteria improved (i.e., smaller BIC) (Supplemental Content Digital Content 1, http://links.lww.com/AA/B399). Although the BIC continued to decrease beyond 9 classes, there were negligible improvements in the clinical characteristics and risk stratification with each additional class identified. Therefore, we focused on the 9-class model for initial analysis. The characteristics of each latent class in the 9-class model (Supplemental Digital Content 2, http://links.lww.com/AA/B399) and a qualitative description of each class (Supplemental Digital Content 3, http://links.lww.com/AA/B399) are available. After preliminary analyses, we determined that there were latent classes with similar adjusted mortality risks that could be combined. Specifically, classes C, D, E, and F (classes are ordered in increasing mortality risk) were combined into 1 class and classes G and H were combined into 1 class. These combinations resulted in fewer latent risk classes that resulted in minimal differences in discrimination (AUC, 0.875 [0.872–0.878] in unadjusted 9-class model versus AUC 0.869 [0.866–0.872] in unadjusted 5-class model). The final sample for analysis contained 5 separate risk classes based on the LCA model.

Back to Top | Article Outline

Baseline Characteristics and Description of the Latent Risk Classes

The baseline characteristics of each latent risk class and general descriptions of each latent risk class are presented in Tables 1 and 2, respectively. (Table 2 describes each latent risk class but is not intended to be used as criteria for classifying patients into specific risk classes.) The first latent risk class (class 1; Table 1) had 56,302 patients (12%) and consisted mainly of young (age <50 years), healthy patients undergoing emergent procedures (65%) (Table 2). The most common procedure for this risk class was appendectomy (59%), followed by cholecystectomy (12%) and colorectal resection (9%). The second risk class (class 2) had 64,100 (14%) patients and consisted mainly of young (age <60 years), morbidly obese females, with 97% having a BMI of >35 kg/m2. The most common procedure for this risk class was gastric bypass (67%), followed by other hernia repair (10%) and cholecystectomy (10%).

Table 1

Table 1

Table 2

Table 2

The third latent risk class (class 3) was the largest class with 213,728 (46%) patients and consisted mainly of a slightly older demographic (age <70 years) with a minor comorbidity burden, mainly hypertension, diabetes mellitus, and cancer. Most procedures were nonemergent (90%). The most common procedure for this risk class was colorectal resection (26%), followed by cholecystectomy (13%) and other hernia repair (12%). The fourth latent risk class (class 4) had 99,169 (21%) patients and consisted of older (age >60 years) patients with a moderate comorbidity burden, including cardiovascular disease, diabetes mellitus, CKD, and cancer. Approximately one third of the procedures were emergent. The most common procedure for this risk class was colorectal resection (32%), followed by cholecystectomy (13%) and appendectomy (12%).

The final latent risk class (class 5) had 32,878 (7%) patients and consisted of older (age >50 years) patients with a severe comorbidity burden, including cardiovascular disease, congestive heart failure, liver disease, CKD requiring dialysis, and pulmonary disease. There were high rates of sepsis (69%) and bleeding disorders (28%), and the procedures were mostly emergent (60%). The most common procedure for this risk class was colorectal resection (29%), followed by exploratory laparotomy (14%) and small bowel resection (11%).

Figure 2

Figure 2

Figure 2 provides a visual representation of the differences in the distribution of comorbidities and risk factors among the latent risk classes. It is clear that class 5 has the highest proportion of patients for most comorbidities and risk factors (Fig. 2, B–D). Although most classes have a similar distribution of BMI values (Fig. 2E), class 2 distinguishes itself because virtually all patients have a BMI of >35 kg/m2.

Back to Top | Article Outline

Latent Risk Classes and 30-Day Mortality Risk

Table 3

Table 3

The overall 30-day mortality among intraabdominal general surgery patients was 2.6% (Table 3). The latent risk classes served as clinically meaningful strata of mortality risk, with the 30-day mortality ranging from 0.06% in class 1 to 22% in class 5, representing a 366-fold difference in mortality between the lowest and highest latent risk classes. Compared with class 4, the RR (95% confidence interval [CI]) for mortality ranged from 0.020 (0.014–0.027) in class 1 to 6.75 (6.49–7.02) in class 5. After adjusting for the procedure category, the adjusted RR (95% CI) ranged from 0.041 (0.029–0.057) in class 1 to 5.42 (5.20–5.65) in class 5, representing a 133-fold difference in the adjusted risk of 30-day mortality between the lowest and highest latent risk classes. Analyses of bootstrapped samples where the assigned latent risk class of an individual patient can vary based on the posterior probabilities demonstrated no material differences between the analytical approaches (Supplemental Digital Content 4, http://links.lww.com/AA/B399).

Back to Top | Article Outline

ASA Physical Status and 30-Day Mortality Risk

Table 4

Table 4

We examined the relationship between the ASA physical status and 30-day mortality. The 30-day mortality ranged from 0.06% in ASA physical status I patients to 56% in ASA physical status V patients (Table 4). The majority of patients were ASA physical status II or ASA physical status III, with each category comprising >40% of the total sample. Compared with ASA physical status III patients, the RR (95% CI) for 30-day mortality ranged from 0.024 (0.016–0.037) in ASA physical status I patients to 24.5 (23.4–25.7) in ASA physical status V patients. After adjusting for procedure, the adjusted RR (95% CI) ranged from 0.039 (0.025–0.061) in ASA physical status I patients to 13.5 (12.8–14.2) in ASA physical status V patients.

Back to Top | Article Outline

Latent Risk Classes and ASA Physical Status

The relationship between the latent risk classes and the ASA physical status was examined (Fig. 3). Most patients in class 1 were ASA physical status I or II, whereas most patients in classes 2, 3, and 4 were ASA physical status II or III. Among class 4 patients, 8.6% were ASA physical status IV. Most patients in class 5 were ASA physical status III and IV, whereas 5.7% were ASA physical status V.

Figure 3

Figure 3

Table 5

Table 5

We examined whether the latent risk classes could further differentiate mortality risk beyond the ASA physical status by conducting analyses stratified by the ASA physical status (Table 5). For all ASA physical status strata, the latent risk classes were significantly associated with 30-day mortality in analyses adjusted for surgical procedure. In addition, for all ASA physical status strata, the AUC of the model with both latent risk class and surgical procedure was significantly greater than the AUC of the model with only the surgical procedure.

Back to Top | Article Outline

Assessing the Improvement in Mortality Prediction When Adding Latent Risk Class

Figure 4

Figure 4

The AUC of the model for 30-day mortality incorporating only the latent risk class was greater than the AUC for the model with only the ASA physical status (Tables 3 and 4; Fig. 4; 0.869 vs 0.847, P < 0.0001). The AUC for the model with the latent risk class and surgical procedure was slightly higher than the AUC for the model with the ASA physical status and surgical procedure (Tables 3 and 4; Fig. 4; 0.895 vs 0.892; P = 0.04). The addition of the latent risk class to a model containing the ASA physical status and procedure resulted in a significant increase in model discrimination (Fig. 4; 0.915 vs 0.892; P < 0.0001).

Back to Top | Article Outline

Calibration Plots and Histograms

Figure 5

Figure 5

Calibration plots for the relative risk regression models for 30-day mortality were plotted by grouping the sample into deciles based on predicted mortality and plotting the predicted versus observed mortality for each group (Fig. 5, A, C, and E). Comparing the fitted linear regression line (dashed line) to the line representing perfect calibration (solid line), the models containing latent risk class and procedure (Fig. 5A), ASA physical status and procedure (Fig. 5C), and latent risk class, ASA physical status, and procedure (Fig. 5E) appear to be well calibrated. The distribution of patients who survived and died within 30 days, within bands of predicted mortality for each model, is displayed adjacent to the calibration plot (Fig. 5, B, D, and F).

Back to Top | Article Outline

Reproducibility of LCA Methodology to Classify Patients

Figure 6

Figure 6

Using the parameter estimates obtained from the original LCA model, we classified patients using the 2011 ACS-NSQIP to determine whether the LCA model could categorize new patients into similar latent risk classes. As in the original analysis, the 9-class model was fit and then reduced to 5 separate risk classes. The distribution of risk factors and comorbidities in the latent risk classes in the new data set was essentially the same as in the original data set, demonstrating that the LCA model parameters could create risk classes with similar characteristics in a new sample (data not shown). In addition, the performance of the regression models evaluating 30-day mortality on latent risk class, procedure, and ASA physical status was nearly identical, with overlapping receiver operating characteristic curves (Fig. 6) and similar AUC values (AUC, 0.915; 95% CI, 0.913–0.917, in original [2005–2010] data set and AUC, 0.917; 95% CI, 0.910–0.923, in new [2011] data set).

Back to Top | Article Outline

DISCUSSION

In a novel application of LCA, we classified patients undergoing intraabdominal general surgery based on their preoperative comorbidities and risk factors into 5 risk classes. These risk classes had clinical characteristics that were meaningfully different from each other and represented distinct categories of patients presenting for surgery. Furthermore, the risk classes stratified patients with regard to 30-day mortality, with a 133-fold difference in mortality between the lowest and highest risk classes, after adjusting for procedure. The addition of the risk classes to models already incorporating the ASA physical status and surgical procedure improved the mortality prediction models as demonstrated by a significant increase in the AUC. Thus, LCA is a viable approach to the analysis of comorbidities and risk factors that provides valuable insights into their relationships with perioperative mortality and also serves to augment the prognostic information contained in the ASA physical status classification.

With conventional approaches to perioperative risk stratification, an individual patient is evaluated and classified into categories based on prespecified criteria. However, LCA is an empirical approach to identify groups of patients with similar characteristics without prespecified constraints, allowing for the recognition of the patterns of comorbidities and risk factors that are actually observed in patients presenting for surgery.

A key advantage to the use of latent variable models is that they provide an effective way to reduce the dimensionality of data, that is, to reduce the complex interrelationships among many variables into a smaller number of factors.29 Indeed, the relationships among the 31 different variables used in our analysis were effectively reduced to a single variable, the latent risk class, which was significantly associated with perioperative mortality. The latent risk class identified 5 comorbidity/risk factor patterns that exist in general surgery patients: (1) young, healthy patients having emergent procedures; (2) young, morbidly obese females; (3) middle-aged patients with hypertension or diabetes mellitus having elective surgery; (4) older patients with more severe cardiovascular disease, including prior coronary revascularizations and strokes, and CKD; and finally (5) older patients with a severe comorbidity burden, including pulmonary disease, liver disease, and sepsis, having emergent surgery. Traditional regression models require complex statistical techniques, such as multiple interaction terms, to adequately account for the interrelationships among all of the comorbidities and risk factors, but this might make the model difficult to interpret.

It could be argued that the ASA physical status is also a latent variable because it is an abstract, unmeasurable variable. The ASA physical status was never intended to measure operative risk,4 but it is highly correlated with clinical risk factors and surgical outcomes and is widely used in models that predict surgical morbidity and mortality.2,3,30 The ASA physical status is determined by the anesthesia provider after a focused history and physical examination and involves the synthesis and interpretation of the data. Ultimately, the ASA physical status is subjective, and different providers may place the same patient into different ASA physical status classes.31,32

In our analysis, we found that the ASA physical status excelled at identifying those at an extremely high risk of perioperative mortality. ASA physical status V patients had a 30-day mortality of 56%, and we were unable to find any other risk stratification measure that was able to identify a subset of patients with a similarly high predicted mortality. However, at 0.5%, these extremely sick patients accounted for only a small fraction of the sample. For most patients, there was considerable variation in mortality risk that was explained in part by the latent risk class. The majority of patients were ASA physical status II or ASA physical status III, with each accounting for approximately 40% of patients. Among ASA physical status II patients, the overall 30-day mortality was 0.30% but, when accounting for latent risk class, ranged from 0.4% to 5.2%. Among ASA physical status III patients, the 30-day mortality was 2.3%, but ranged from 0.38% to 11% when accounting for latent risk class. Finally, among ASA physical status IV patients, the overall 30-day mortality was 18% but ranged from 0.44% to 29% when accounting for latent risk class.

We also identified a group of patients with a mortality risk lower than that predicted by the ASA physical status. Latent risk class 2 was comprised mainly of young, female patients who were morbidly obese, and not surprisingly, two thirds were undergoing gastric bypass procedures. More than half of the patients were thought to have severe systemic disease or worse, with 55% classified as ASA physical status III or higher. However, the overall 30-day mortality of this risk class was only 0.12%, and even the ASA physical status III and ASA physical status IV patients had a low mortality risk of 0.15% and 0.47%, respectively, implying that the ASA physical status was not the best predictor of mortality in this group. This finding could represent the obesity paradox,33 the protective effects of sex hormones in women,34 or the synergistic effects of both. Alternatively, this may simply reflect the fact that the main procedure, gastric bypass, has an inherently low mortality risk,35 and these effects need to be further clarified.

From a model performance perspective, the latent risk class was an excellent predictor of 30-day mortality, with an AUC of 0.869 alone and 0.895 with the surgical procedure. The ASA physical status was equally as effective for predicting mortality, with an AUC of 0.847 alone and 0.892 with the surgical procedure. Because the ASA physical status is a well-established risk stratification measure, we measured the incremental improvement in risk prediction when adding the latent risk class to the ASA physical status (and surgical procedure). With all 3 variables, the AUC increased to 0.915. The AUC of the model with the ASA physical status and surgical procedure was already high, so it would be difficult to demonstrate large increases in discrimination with the addition of the latent risk class. Using the net reclassification improvement and integrated discrimination improvement,36 metrics designed to quantify the impact of an additional variable to a model, we found that the main improvement with the addition of the latent risk class involved a more accurate prediction of mortality events among those dying within 30 days of surgery (data not shown).

Because this is a novel application of LCA to perioperative mortality risk stratification, the optimal use of the latent risk classes remains to be determined. The ASA physical status has withstood the test of time with excellent discrimination in perioperative mortality stratification models,37 and it may not be possible for alternate classification systems to substantially exceed its performance characteristics. We posit that the ASA physical status represents the practitioner’s “gut” feeling after evaluating all of the available information, and although gut feelings may be reliable in some instances,38 we have demonstrated that there are significant variations in mortality even within a given ASA physical status class. Our results suggest that a strategy combining the ASA physical status with a complementary system such as latent risk classes may prove beneficial, especially among patients who are not at either extreme (extremely low or extremely high) of mortality risk. An additional area of investigation will be to assess the relationships between the latent risk classes and perioperative morbidities because prior studies demonstrate differences in the predictive ability of preoperative data with respect to mortality and morbidity.37 The use of other latent variable methods may reveal additional important associations between perioperative risk factors and clinical outcomes, and we plan to use techniques such as factor analysis and structural equation modeling to further explore these relationships.

A major strength of our analysis is that all of the available data were used in the LCA model. We did not have to make decisions on which variables to include or omit based on statistical significance or other criteria. In addition, we included all patients and had no exclusion criteria other than procedure type and inpatient surgery. Thus, we have a model that can be applied to all patients presenting for intraabdominal general surgery. We also included variables that might not necessarily be used in other risk stratification tools, such as age. Age is not an explicit component of the ASA physical status, but providers may implicitly account for it when making their assessment, especially at the extremes of age.39

There are some limitations to our LCA model of perioperative risk. We identified classes of patients with similar characteristics, and based on these prevailing characteristics, a qualitative description of each class was defined. However, there is still some variation within each class, and not all individuals assigned to a particular class are identical with respect to their comorbidities and risk factors. As such, there is variability among individuals in terms of their likelihood to be in different classes, and there is a potential for misclassification.

Another limitation of our analysis is that in determining the optimal number of latent classes, we settled on the 9-class model. However, the empirical model comparison measure, BIC, indicated that a 10-class model fit better and we did not test models above 10 classes because they were computationally prohibitive. The improvement in fit as the model becomes more complicated is driven by the very large sample size that can support complicated models. But despite the possibility that further increasing the numbers of latent classes might have better fit the data, a parsimonious clinically interpretable solution was also highly desirable. We chose to analyze a broad sample of intraabdominal general surgery patients, but there is some degree of heterogeneity in these procedures, and separate LCA models for each procedure might have yielded different results. Because it relies on complex statistical modeling, it is not a simple task to accurately determine any given patient’s latent risk class, but this could be easily implemented as a function in an electronic medical record or online calculator.

We used LCA based on our belief that there is an underlying latent structure to the variables analyzed (i.e., the pattern of comorbidities and risk factors are explained by an unmeasured latent variable, the comorbid state). However, there are other classification methods, such as cluster analysis,40 and the classification of a sample may vary depending on the methodology used.41 It is possible that a different classification technique may offer different insights into the data, and this may need to be explored further.

In conclusion, we have demonstrated that LCA is a novel and effective method of categorizing patients presenting for intraabdominal general surgery based on their comorbidities and risk factors. The resulting latent risk classes were significantly associated with 30-day mortality and demonstrated significant improvements to models incorporating the ASA physical status. This underscores the potential value of latent variable methodologies to improve our understanding of the relationships between comorbidities and perioperative mortality.

Appendix 1

Appendix 1

Appendix 2

Appendix 2

Back to Top | Article Outline

DISCLOSURES

Name: Minjae Kim, MD, MS.

Contribution: This author was involved in study design, conduct of the study, data analysis, manuscript preparation, and is the archival author.

Attestation: Minjae Kim approved the final manuscript and attests to the integrity of the original data and analysis reported in this manuscript.

Name: Melanie M. Wall, PhD.

Contribution: This author was involved in study design, data analysis, and manuscript preparation.

Attestation: Melanie M. Wall approved the final manuscript and attests to the integrity of the original data and analysis reported in this manuscript.

Name: Guohua Li, MD, DrPH.

Contribution: This author was involved in study design, data analysis, and manuscript preparation.

Attestation: Guohua Li approved the final manuscript and attests to the integrity of the original data and analysis reported in this manuscript.

This manuscript was handled by: Franklin Dexter, MD, PhD.

Back to Top | Article Outline

FOOTNOTES

a Available at: http://www.asahq.org/resources/clinical-information/asa-physical-status-classification-system. Accessed February 23, 2015.
Cited Here...

b Available at: http://www.john-uebersax.com/stat/faq.htm. Accessed June 20, 2013.
Cited Here...

c Available at: https://methodology.psu.edu/ra/lca. Accessed December 30, 2015.
Cited Here...

d Available at: http://www.ats.ucla.edu/stat/mplus/seminars/lca/. Accessed December 30, 2015.
Cited Here...

e Available at: http://www.john-uebersax.com/stat/soft.htm. Accessed December 30, 2015.
Cited Here...

f The American College of Surgeons National Surgical Quality Improvement Program (ACS-NSQIP) and the hospitals participating in the ACS-NSQIP are the source of the data used herein; they have not verified and are not responsible for the statistical validity of the data analysis or the conclusions derived by the authors.
Cited Here...

g Available at: http://www.hcup-us.ahrq.gov/toolssoftware/ccs_svcsproc/ccssvcproc.jsp. Accessed June 15, 2012.
Cited Here...

h Available at: https://www.facs.org/~/media/files/quality%20programs/nsqip/ug10.ashx. Accessed November 3, 2015.
Cited Here...

Back to Top | Article Outline

REFERENCES

1. Vacanti CJ, VanHouten RJ, Hill RCA statistical analysis of the relationship of physical status to postoperative mortality in 68,388 cases.Anesth Analg197049564–6
2. Wolters U, Wolf T, Stützer H, Schröder TASA classification and perioperative variables as predictors of postoperative outcome.Br J Anaesth199677217–22
3. Glance LG, Lustik SJ, Hannan EL, Osler TM, Mukamel DB, Qian F, Dick AWThe Surgical Mortality Probability Model: derivation and validation of a simple risk prediction rule for noncardiac surgery.Ann Surg2012255696–702
4. Saklad MGrading of patients for surgical procedures.Anesthesiology19472281–4
5. Ouellette JR, Small DG, Termuhlen PMEvaluation of Charlson-Age Comorbidity Index as predictor of morbidity and mortality in patients with colorectal carcinoma.J Gastrointest Surg200481061–7
6. Hines RB, Chatla C, Bumpers HL, Waterbor JW, McGwin G Jr, Funkhouser E, Coffey CS, Posey J, Manne UPredictive capacity of three comorbidity indices in estimating mortality after surgery for colon cancer.J Clin Oncol2009274339–45
7. Elixhauser A, Steiner C, Harris DR, Coffey RMComorbidity measures for use with administrative data.Med Care1998368–27
8. Lee TA, Shields AE, Vogeli C, Gibson TB, Woong-Sohn M, Marder WD, Blumenthal D, Weiss KBMortality rate in veterans with multiple chronic conditions.J Gen Intern Med200722Suppl 3403–7
9. Bollen KALatent variables in psychology and the social sciences.Annu Rev Psychol200253605–34
10. Rabe-Hesketh S, Skrondal AClassical latent variable models for medical research.Stat Methods Med Res2008175–32
11. Anand P, Krishnakumar J, Tran NBMeasuring welfare: latent variable models for happiness and capabilities in the presence of unobservable heterogeneity.J Public Econ201195205–15
12. Conway ARA, Cowan N, Bunting MF, Therriault DJ, Minkoff SRBA latent variable analysis of working memory capacity, short-term memory capacity, processing speed, and general fluid intelligence.Intelligence200230163–83
13. Everitt BAn Introduction to Latent Variable Models1984London; New YorkChapman and Hall
14. Tsai J, Rosenheck RAPsychiatric comorbidity among adults with schizophrenia: a latent class analysis.Psychiatry Res201321016–20
15. Galatzer-Levy IR, Nickerson A, Litz BT, Marmar CRPatterns of lifetime PTSD comorbidity: a latent class analysis.Depress Anxiety201330489–96
16. Rindskopf D, Rindskopf WThe value of latent class analysis in medical diagnosis.Stat Med1986521–7
17. Neuhaus V, Ring DCLatent class analysis.J Hand Surg Am2013381018–20
18. Formann AK, Kohlmann TLatent class analysis in medical research.Stat Methods Med Res19965179–211
19. Fink AS, Campbell DA Jr, Mentzer RM Jr, Henderson WG, Daley J, Bannister J, Hur K, Khuri SFThe National Surgical Quality Improvement Program in non-veterans administration hospitals: initial demonstration of feasibility.Ann Surg2002236344–53
20. Hua M, Brady J, Li GThe epidemiology of upper airway injury in patients undergoing major surgical procedures.Anesth Analg2012114148–51
21. Kim M, Brady JE, Li GVariations in the risk of acute kidney injury across intraabdominal surgery procedures.Anesth Analg20141191121–32
22. World Health OrganizationObesity: Preventing and Managing the Global Epidemic: Report of a WHO Consultation2000Geneva, SwitzerlandWorld Health Organization
23. Levey AS, Stevens LA, Schmid CH, Zhang YL, Castro AF III, Feldman HI, Kusek JW, Eggers P, Van Lente F, Greene T, Coresh JCKD-EPI (Chronic Kidney Disease Epidemiology Collaboration)A new equation to estimate glomerular filtration rate.Ann Intern Med2009150604–12
24. National Kidney FoundationK/DOQI clinical practice guidelines for chronic kidney disease: evaluation, classification, and stratification.Am J Kidney Dis200239S1–266
25. Schwarz GEstimating the dimension of a model.Ann Statist19786461–4
26. Nylund KL, Asparouhov T, Muthén BODeciding on the number of classes in latent class analysis and growth mixture modeling: a Monte Carlo simulation study.Struct Equ Modeling200714535–69
27. Enders CK, Bandalos DLThe relative performance of full information maximum likelihood estimation for missing data in structural equation models.Struct Equ Modeling20018430–57
28. Spiegelman D, Hertzmark EEasy SAS calculations for risk or prevalence ratios and differences.Am J Epidemiol2005162199–200
29. Bartholomew DJ, Knott M, Moustaki ILatent Variable Models and Factor Analysis: A Unified Approach20113rd edChichester, West Sussex, United KingdomWiley
30. Davenport DL, Bowe EA, Henderson WG, Khuri SF, Mentzer RM JrNational Surgical Quality Improvement Program (NSQIP) risk factors can be used to validate American Society of Anesthesiologists Physical Status Classification (ASA PS) levels.Ann Surg2006243636–41
31. Mak PH, Campbell RC, Irwin MGAmerican Society of AnesthesiologistsThe ASA Physical Status Classification: inter-observer consistency. American Society of Anesthesiologists.Anaesth Intensive Care200230633–40
32. Haynes SR, Lawler PGAn assessment of the consistency of ASA physical status classification allocation.Anaesthesia199550195–9
33. Gruberg L, Weissman NJ, Waksman R, Fuchs S, Deible R, Pinnow EE, Ahmed LM, Kent KM, Pichard AD, Suddath WO, Satler LF, Lindsay J JrThe impact of obesity on the short-term and long-term outcomes after percutaneous coronary intervention: the obesity paradox?J Am Coll Cardiol200239578–84
34. Horton JW, White DJ, Maass DLGender-related differences in myocardial inflammatory and contractile responses to major burn trauma.Am J Physiol Heart Circ Physiol2004286H202–13
35. Buchwald H, Avidor Y, Braunwald E, Jensen MD, Pories W, Fahrbach K, Schoelles KBariatric surgery: a systematic review and meta-analysis.JAMA20042921724–37
36. Pencina MJ, D’Agostino RB Sr, D’Agostino RB Jr, Vasan RSEvaluating the added predictive ability of a new marker: from area under the ROC curve to reclassification and beyond.Stat Med200827157–72
37. Moonesinghe SR, Mythen MG, Das P, Rowan KM, Grocott MPRisk stratification tools for predicting morbidity and mortality in adult patients undergoing major surgery: qualitative systematic review.Anesthesiology2013119959–81
38. Markus PM, Martell J, Leister I, Horstmann O, Brinker J, Becker HPredicting postoperative morbidity by clinical assessment.Br J Surg200592101–6
39. Fitz-Henry JThe ASA classification and peri-operative risk.Ann R Coll Surg Engl201193185–7
40. Everitt BCluster Analysis20115th ed.Chichester, West Sussex, United KingdomWiley
41. Eshghi A, Haughton D, Legrand P, Skaletsky M, Woolford SIdentifying groups: a comparison of methodologies.J Data Sci20119271–91

Supplemental Digital Content

Back to Top | Article Outline
© 2016 International Anesthesia Research Society