Abstract
Objective:
To develop a claims-based model to predict persistent high-dose opioid use amongst patients undergoing total knee replacement (TKR).
Methods:
Using Medicare claims (2010–2014), we identified patients ≥65 years who underwent TKR with no history of high-dose opioid use (>25 mean morphine equivalents (MME)/day) in the year prior. We used group-based trajectory modeling to identify distinct opioid use patterns. The primary outcome was persistent high-dose opioid use in the year after TKR. We split the data into training (2010–2013) and test (2014) sets and used logistic regression with least absolute shrinkage and selection operator (LASSO) regularization utilizing a total of 83 pre-operative patient characteristics as candidate predictors. A reduced model with ten pre-specified variables which included demographics, opioid use and medication history was also considered.
Results:
The final study cohort included 142,089 patients who underwent TKR. The group-based trajectory model identified 4 distinct trajectories of opioid use (Group 1- short-term, low-dose, Group 2- moderate-duration, low-dose, Group 3- moderate-duration, high-dose, and Group 4-persistent high-dose). The model predicting persistent high-dose opioid use achieved high discrimination (area under the receiver operating characteristic curve (AUC) of 0.85; 95% CI, 0.84–0.86)) in the test set. The reduced model with ten predictors performed equally well (AUC=0.84; 95% CI, 0.84–0.85).
Conclusions:
In this cohort of older patients, 10.6% became persistent high dose (mean=22.4 MME/day) opioid users after TKR. Our model with 10 readily available clinical factors may help identify patients at high risk of future adverse outcomes from persistent opioid use after TKR.
Introduction
The economic burden of prescription opioid overdose, abuse, and dependence is estimated to be $78.5 billion each year in the United States. 1 An estimated 2.0 million people in the United States had opioid use disorder (defined in the DSM-5 as a problematic pattern of opioid use leading to clinically significant impairment or distress)2 associated with prescription opioids in 2015 and nearly half of all opioid related deaths involved a prescription opioid.3 4
Patients undergoing major surgical procedures such as total knee replacements (TKR) are often prescribed opioids before and/or after surgery for pain relief. A recent study revealed that 87.1 % of patients filled at least 1 prescription for opioids in the year prior to hip or knee arthroplasty in a commercially insured population.5 6 In a large cohort of older Medicare enrollees with osteoarthritis, 58.3% had used opioids at least once in the year prior to TKR, and 7.2% had continuous opioid use, defined by a dispensing for opioid at least once every month for 12 months before the surgical procedure.7 Some studies suggest that use of pre- or peri-operative opioids increases risk for persistent opioid use and opioid dependence following the surgery.8 9 Patients who were continuous users of opioids prior to surgery were found to have poorer surgical outcomes after surgery and nearly a 5-fold increased risk for opioid overdose compared to those who did not use any opioids prior to surgery. 7
Few studies have characterized the longitudinal patterns of opioid use after TKR5, and those that do have not studied in detail the dynamic patterns of use over time. Group-based trajectories have been used to model complex longitudinal outcomes or behaviors such as healthcare spending10, post-operative pain11 and medication adherence to chronic medications12 13, and may aid in characterizing longitudinal patterns of opioid use over time.14
The objectives of our study were to: 1) characterize the dynamic patterns of opioid use of patients who underwent TKR in the year following the surgery using group-based trajectory modeling to classify patients with persistent high-dose opioid use; and 2) develop a prognostic clinical prediction model to identify persistent high-dose opioid users after TKR using pre-operative patient characteristics.
Patients and Methods
Data Source and study population
Using Medicare Parts A (inpatient), B (outpatient) and D (prescription) claims (2010–2014) and using a cohort design, we defined cohort eligibility as patients aged ≥65 years who underwent a TKR (e-Table 4) and were continuously enrolled in Medicare for ≥360 days prior to their TKR and for a minimum of 30 days after TKR. To prevent inclusion of patients with bilateral TKRs, patients with two codes for TKR on the same day or with a history of TKR in the 360 days prior were not included. We defined two exclusion criteria; 1) patients with any prior history of cancer as these patients may be more likely to have an indication for persistent opioid use, and 2) patients with high-dose opioid use at baseline (>25 Morphine Milligram Equivalents (MME)/day) as previous use of high-dose opioids itself is a strong predictor of future high-dose opioids.15 16 Opioid use was assessed in the 360 days after TKR in 30 day intervals. To calculate MME/day, we utilized the “Opioid NDC and Oral MME Conversion File” provided by the CDC’s National Center for Injury Prevention and Control17 that contains opioid NDCs with their linked oral morphine milligram equivalent (MME) conversion factors (e-Table 3). If patients filled multiple prescriptions of the same drug on the same day, we only considered the fill with a higher days’ supply. However, prescriptions filled for different opioid medications on the same day or in combination with other medications (such as NSAIDs) contributed to the MME calculation.
A signed data use agreement with the Center for Medicare and Medicaid services was available, and the Brigham and Women’s Hospital Institutional Review Board approved the protocol for this study. The individual patient–level data will not be made available to other researchers for purposes of reproducing the results or replicating the procedure to protect patient privacy. Programming codes will be made available upon request on publication of this study to enable other researchers to implement the model proposed in our study.
Group-based trajectory modeling
A trajectory model estimates several regression models simultaneously through maximization of a likelihood function that combines the information from all models.14 Within each opioid use group, usage patterns are modeled as a smooth function of time. The output of a group-based trajectory includes estimated probabilities of group membership for each individual and an estimated trajectory curve over time for each group.14 Group based trajectory models were built using opioid filling patterns in the 360 days after TKR. We modeled opioid use as continuous variable (MME) in every 30-day interval using Proc Traj in SAS 9.4 and specifying a censored normal model. Multiple models were developed varying the number of groups from 2 to 6; the 4-group model was selected as the final model based on assessment of model fit. This included a combination of factors including the Bayesian Information Criterion (BIC), the predicted probability of group membership (typically >0.9)14 and having an adequate number of patients in each trajectory group.
Outcomes – Defining persistent high-dose opioid users
The group-based trajectory model classified patients into 4 distinct trajectories (Group 1- short-term, low-dose, Group 2- moderate-duration, low-dose, Group 3- moderate-duration, high-dose, and Group 4-persistent high-dose) of opioid use in the year after TKR. The primary outcome was persistent high-dose opioid use defined as patients in trajectory Group 4. Patients in Group 4 had the longest average duration (164.4 days) of opioid use and high average dose (22.4 MME/day) compared to patients in trajectory Group 1 (14.3 days and 20.9 MME/day), Group 2 (44.9 days and 14.4 MME/day), and Group 3 (51.1 days and 38.8 MME/day) (Figure-1).
Figure 1 – Trajectories of opioid filling patterns after TKR.
Footnote: MME= mean morphine equivalent milligrams, SD=standard deviation
Development of prediction model
We used logistic regression to predict membership in trajectories with persistent high-dose opioid use (Groups 4) vs other groups (Groups 1, 2 and 3) as a binary outcome utilizing all candidate predictors and implemented the least absolute shrinkage and selection operator (LASSO) regression for variable selection using the GLMNET package in R18. We performed 10-fold cross-validation to select the value of the penalty parameter in a way that minimized the mean cross-validated error. The study cohort was split into a training set (2010–2013; 71.7% of the sample) and a test set (2014; 28.3% of the sample) based on the year of cohort entry. We used the regression coefficients for each of the selected predictor variables estimated in the training set to predict persistent high-dose vs. other groups in the test set.
Predictors
Based on Medicare medical, procedure or pharmacy claims, we defined a total of 83 investigator-specified candidate predictors in the 360 days prior to the date of TKR (the index date). These included demographics (such as age, gender, and race), comorbidities (such as substance use disorders, depression, arthritis), co-medications (such as baseline opioid use, benzodiazepines, and antidepressants), healthcare utilization variables (such as emergency room visits, hospitalizations, and physician office visits), comorbidity index19, and markers of frailty captured using a validated claims-based frailty index.20 We also considered a reduced model with only 10 pre-specified which included demographics, history of opioid use and other substance use, and prior medication history. The rationale for these 10 variables was to include variables that are readily available for clinical use but also reliably captured in claims data (such as medication use) that may enhance the external validity of the model in other settings.
Performance metrics and diagnostics
Predictive performance of the model was assessed by calculating the area under the curve (AUC) of the receiver operating characteristic (ROC) curve with values close to 0.5 indicating an uninformative model and values close to 1 indicating perfect prediction. Since the AUC of the ROC curve can be too optimistic, we also calculated the area under the precision-recall curve (AUPRC), that factors in the outcome prevalence and visually plots the tradeoff between positive predictive value (PPV) and sensitivity.21 In contrast to the AUC of the ROC curve, the lower bound of the AUPRC is determined by the outcome prevalence (approximately 0.10 in our study given 10% outcome prevalence) which indicates an uninformative model and values close to 1 indicating perfect prediction. Plots of ROC and PRC were generated in both the training and test data. Finally, calibration plots were produced to assess if the predicted probabilities match the actual probabilities by visual inspection and reporting the intercept and slope. Intercept values close to zero and slope values close to 1 indicate perfect prediction.22 23 24
Results
We identified 142,089 patients aged ≥65 years who underwent TKR and met all eligibility criteria (eFigure-1). The 4-group trajectory model identified clusters of patients (Figure-1) with clinically meaningful discrimination between patients’ opioid use. Patient characteristics between patients in Group 4 (long-term, high-dose opioid use) vs. Group 1 (short-term, low-dose opioid use) differed greatly (Table-1). Patients in Group 4 vs. Group 1 were more likely to be younger, female, have higher baseline opioid use (9.0 MME/day vs 0.8 MME/day, respectively) in the year prior to TKR, more likely to use tobacco/alcohol and had a greater burden of comorbidities such as depression, anxiety disorders, back pain and used multiple medications. Patients in Group 4 were also more likely to be frail as assessed by the claims-based frailty index20 and were also more likely to use healthcare services in general such as number of ER visits and physician office visits.
Table-1:
Selected baseline patient characteristics by opioid use trajectory groups
| Group 1 (Short-term, Low-Dose) | Group 2 (Moderate duration, Low-Dose) | Group 3 (Moderate duration, High-Dose) | Group 4 (Persistent High-Dose) | |
|---|---|---|---|---|
|
| ||||
| N | 80,801 | 38,224 | 8,020 | 15,044 |
| Age (Mean, SD), years | 73.09 (5.35) | 72.68 (5.23) | 70.64 (4.09) | 72.03 (5.09) |
| Male (%) | 30025 (37.2) | 11887 (31.1) | 3018 (37.6) | 4089 (27.2) |
| White race (%) | 74686 (92.4) | 33757 (88.3) | 7445 (92.8) | 13228 (87.9) |
| Combined Comorbidity index19 (Mean, SD) | 0.50 (1.64) | 0.71 (1.78) | 0.45 (1.51) | 0.95 (1.94) |
| Frailty categories20 | ||||
| Mild (<0.13) (%) | 21110 (26.1) | 7000 (18.3) | 2136 (26.6) | 1756 (11.7) |
| Moderate (0.13-<0.16) (%) | 32015 (39.6) | 14032 (36.7) | 3027 (37.7) | 4725 (31.4) |
| Severe (>=0.16) (%) | 27676 (34.3) | 17192 (45.0) | 2857 (35.6) | 8563 (56.9) |
| Baseline MME/day (Mean, SD) | 0.77 (1.95) | 2.18 (3.54) | 2.27 (4.23) | 9.02 (7.00) |
| Tobacco use (%) | 8230 (10.2) | 4401 (11.5) | 1082 (13.5) | 2279 (15.1) |
| Alcohol abuse (%) | 482 (0.6) | 277 (0.7) | 71 (0.9) | 149 (1.0) |
| Anxiety (%) | 6112 (7.6) | 3927 (10.3) | 837 (10.4) | 2351 (15.6) |
| Falls (%) | 2569 (3.2) | 1600 (4.2) | 243 (3.0) | 904 (6.0) |
| Back pain (%) | 30837 (38.2) | 18019 (47.1) | 3430 (42.8) | 8799 (58.5) |
| Depression (%) | 8051 (10.0) | 5164 (13.5) | 1083 (13.5) | 2933 (19.5) |
| Diabetes (%) | 21488 (26.6) | 12173 (31.8) | 2158 (26.9) | 5131 (34.1) |
| Drug abuse (%) | 39 (0.0) | 37 (0.1) | 11 (0.1) | 54 (0.4) |
| Number of ER Visits (Mean, SD) | 0.25 (0.63) | 0.35 (0.80) | 0.29 (0.78) | 0.47 (1.01) |
| Number of drugs (Mean, SD) | 8.46 (4.86) | 10.60 (5.39) | 9.66 (5.20) | 12.77 (5.83) |
| Number of office visits (Mean, SD) | 10.39 (6.25) | 11.94 (7.03) | 11.09 (6.51) | 13.18 (7.75) |
MME: Mean morphine equivalent milligram; ER: Emergency Room, SD: standard deviation
Using logistic regression and LASSO, we predicted the probability of persistent high-dose opioid use (N=11,607) in the training data (N=101,810) for an AUC of 0.88 (95% CI, 0.88–0.89). The AUC in the test data (N=40,279) predicting high-dose opioid use (N=3,437) was 0.85 (95% CI, 0.84–0.86) (Figure-2, eTable-1). The AUPRC was 0.59 and 0.45, respectively in the training and test data (Figure-2). Visual inspection of the calibration plot suggested that the model tended to overestimate the predicted probability of persistent high-dose opioid use when the actual probabilities were greater than 40% (Calibration slope=0.82). (eFigure-2). The final model with the lowest mean cross-validated error selected 22 predictors (eFigure-3). Figure-3 shows the variables selected in the full model and their coefficients predicting persistent high-dose opioid use. The strongest positive predictor of persistent high-dose opioid use was baseline opioid use. The other variables selected in the final model included demographics (such as age, region and race), medication use (such as use of benzodiazepines, anxiolytic medications, NSAIDs, and antidepressants), substance use and comorbidities (such as migraine, anxiety, and back pain)
Figure 2 –
Comparison of receiver operating characteristic (ROC) and precision-recall (PRC) curves in the full model (Group 4 vs. Group 1,2 and 3)
Figure 3 –
Variables selected in the full model and their coefficients predicting persistent high-dose opioid use.
A reduced model with only ten investigator specified predictors which included demographics (age, sex and race), history of substance abuse (opioids, alcohol and tobacco) and medication use (benzodiazepines, anxiolytics, antidepressants, anticonvulsants and NSAIDs), also showed comparable predictive ability in terms of discrimination (AUC=0.84; 95% CI, 0.84–0.85), precision (AUPRC=0.45) and calibration (Calibration slope=0.79) (eTable-4,5). The strongest predictors of persistent high-dose opioid use was baseline opioid use (modeled as a continuous variable) (Odds Ratio (OR)=1.33; 95%CI, 1.33–1.34), black race (OR=1.50; 95% CI, 1.36–1.66) and history of benzodiazepine use (OR=1.44; 95% CI, 1.32–1.57). (Table-2). The coefficients for the final predictive model are reported in e-Table 2 and an online risk calculator will be publicly available on our website (https://www.bwhprime.org/).
Table 2:
Predictors of trajectories of persistent high-dose opioid use in the reduced model in the training data
| Variable | Multivariable odds ratio (95% CI) Persistent high dose (Group 4) vs. Other groups (Group 1,2,3) | P-value |
|---|---|---|
|
| ||
| Age (in years) | 0.98 (0.98–0.99) | <0.001 |
| Females (ref=males) | 0.83 (0.79–0.88) | <0.001 |
| Race | ||
| Other | 1.15 (1.04–1.28) | 0.008 |
| Black | 1.50 (1.36–1.66) | <0.001 |
| White (ref.) | 1.00 | |
| Baseline opioid use (MME/day) | 1.33 (1.33–1.34) | <0.001 |
| Substance use (yes/no) | 1.14 (1.06–1.22) | <0.001 |
| Benzodiazepine use (yes/no) | 1.44 (1.32–1.57) | <0.001 |
| Anxiolytic use (yes/no) | 1.27 (1.19–1.36) | <0.001 |
| Anticonvulsant use (yes/no) | 1.25 (1.17–1.33) | <0.001 |
| Antidepressant use (yes/no) | 1.16 (1.10–1.23) | <0.001 |
| NSAID use (yes/no) | 1.21 (1.15–1.27) | <0.001 |
CI= confidence interval, MME= mean morphine equivalent milligrams, NSAID= non-steroidal anti-inflammatory drug
Discussion
In this large cohort of older Medicare enrollees who underwent TKR (mean age=72.7 years), group-based trajectory modeling identified groups of patients with four distinct patterns of opioid use in the year following the surgery. We classified one of these groups comprising 10.6% of the population as persistent high-dose users (average dose of 22.4 MME/day for an average duration of 164.4 days) and developed a prediction model using only pre-operative patient characteristics to predict membership in the empirically identified trajectory group which showed excellent predictive performance. Using a reduced model with only ten investigator specified predictors that would be more readily available in clinical settings also resulted in comparable predictive performance.
The key predictors of persistent high-dose opioid use were baseline use of opioids, benzodiazepines, anxiolytics, antidepressants and diagnosis or treatments for chronic painful conditions. Numerous prior studies5 7 25 15 have identified baseline opioid use as an extremely strong predictor of future opioid use and this finding was also observed in our study, even after excluding patients with high dose opioid use at baseline. Although the final model selected 22 predictors, only minimal gains were observed beyond adding baseline opioid use with the most regularized model only selecting 13 predictors. Black race (compared to white individuals) was the second strongest predictor identified and is consistent with recent trends suggesting large increases in rates of opioid overdose among older black men and women.26 27
Other influential predictors were use of benzodiazepines that have been associated with higher rates of opioid dependence and overdose in prior studies5 28 and the presence of chronic painful conditions (the majority of patients in the cohort (99.5%) had a diagnosis of osteoarthritis or rheumatoid arthritis at baseline) requiring non-opioid pain medications such as NSAIDs or gabapentin that have also been identified in other studies as a predictor of future persistent opioid use5 25.
The predictors in the reduced model focused mainly on variables that are well-captured in claims such as demographics and prior medication use (including opioid use) and thus may be more generalizable when used in other settings. Using a simple algorithm to pre-operatively identify patients at high risk of persistent opioid use, may help providers exercise appropriate caution before prescribing opioids after surgery. The predictive performance of our model was comparable to other studies that have aimed to predict prolonged opioid use after orthopedic surgeries.15 25 The tradeoff for gains in positive predictive value to loss of sensitivity may be a worthwhile compromise, given the scale of the opioid epidemic in the US. A greater emphasis on perioperative interventions to improve early mobilization (such as using an adductor canal block compared to femoral nerve block)29 30 and manage post-operative pain using non-opioid pain management strategies after TKR such as use of NSAIDs, gabapentin31, local infiltrating analgesia32 or knee braces32 may be warranted for patients at high risk of future opioid dependence.
Limitations
This study includes older patients enrolled in Medicare and thus, our findings may not be generalizable to other populations. Patients who did not have a minimum of 30 days of follow up after TKR were excluded. Most of these patients (90.2%) were admitted to a nursing home where it is no longer possible to track their prescription claims, however, the demographics of patients excluded were largely similar to those in the final cohort thus minimizing the potential for bias. While the trajectory model empirically identified a group of patients with persistent high-dose opioid use, it remains possible that such use is specific to our study population and thus our prediction model may be prone to overfitting which highlights the need for future work to externally validate our model using alternate data sources before it can be translated into practice. Another limitation is that we are not aware of the specific indication for opioid use although we tried to minimize this issue by excluding patients with cancer and high-dose opioid use prior to surgery. Finally, opioid use is captured in claims data based on filled prescriptions rather than actual consumption which may be susceptible to some misclassification. However, this method of assessing exposure is generally superior to physician prescribing records and patient self-reporting.33 34
Conclusion
In this cohort of 142,089 older patients with no history of cancer or high-dose opioid use at baseline, 10.6% became persistent high dose (mean: 22.4 MME/day) opioid users during the year after TKR. Our prediction model developed using Medicare claims with 10 readily available clinical factors may help identify patients at high risk of future adverse outcomes from persistent opioid use and dependence after TKR.
Supplementary Material
Significance and Innovations.
Using group-based trajectory modeling, our study identified 4 distinct patterns of opioid use amongst Medicare patients who underwent a total knee replacement (TKR) in the year after their surgery.
10.6% of these patients became persistent high dose (mean=22.4 MME/day) opioid users in the year after their TKR.
Our model with 10 readily available pre-operative clinical factors achieved excellent predictive performance (AUC=0.84) and may help identify patients at high risk of future adverse outcomes from persistent opioid use after TKR.
Acknowledgments
Funding/Support:
This study was supported by a grant from the National Institutes of Health and National Institute of Arthritis and Musculoskeletal and Skin Diseases (R01AR069557).
Role of the Funder/Sponsor:
The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.
Conflict of Interest Disclosures:
Dr. Desai reported receiving grants from Bayer, Novartis, and Vertex Pharmaceuticals outside the submitted work.
Dr Katz reported receiving grants from Samumed and Flexion Therapeutics outside the submitted work.
Dr Lee reported receiving grants from Pfizer outside the submitted work, owning stock in Cigna-Express Scripts, and serving as an advisory board member for Eli Lilly.
Dr Kim reported receiving grants from AbbVie, Bristol-Myers Squibb, Pfizer, and Roche (paid to Brigham and Women’s Hospital) outside the submitted work.
Dr. Solomon reported receiving salary support from grants to Brigham and Women’s Hospital from Abbvie, Amgen, Corrona, Genentech, Janssen, and Pfizer.
No other disclosures were reported for Gopalakrishnan, Jin and Lii.
References
- 1.Florence CS, Zhou C, Luo F, Xu L. The Economic Burden of Prescription Opioid Overdose, Abuse, and Dependence in the United States, 2013. Med Care. 2016;54(10):901–906. doi: 10.1097/MLR.0000000000000625 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.American Psychiatric Association: Diagnostic and Statistical Manual of Mental Disorders. 5 th ed. Arlington: American Psychiatric Association; 2013. [Google Scholar]
- 3.Huang B, Dawson DA, Stinson FS, Hasin DS, Ruan WJ, Saha TD, Smith SM, Goldstein RB, Grant BF. Prevalence, Correlates, and Comorbidity of Nonmedical Prescription Drug Use and Drug Use Disorders in the United States: Results of the National Epidemiologic Survey on Alcohol and Related Conditions. J Clin Psychiatry. 2006;67(07):1062–1073. doi: 10.4088/JCP.v67n0708 [DOI] [PubMed] [Google Scholar]
- 4.Guy GP. Vital Signs: Changes in Opioid Prescribing in the United States, 2006–2015. MMWR Morb Mortal Wkly Rep. 2017;66. doi: 10.15585/mmwr.mm6626a4 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Kim SC, Choudhry N, Franklin JM, Bykov K, Eikermann M, Lii J, Fischer MA, Bateman BT. Patterns and predictors of persistent opioid use following hip or knee arthroplasty. Osteoarthritis Cartilage. 2017;25(9):1399–1406. doi: 10.1016/j.joca.2017.04.002 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Wright EA, Katz JN, Abrams S, Solomon DH, Losina E. Trends in prescription of opioids from 2003–2009 in persons with knee osteoarthritis. Arthritis Care Res. 2014;66(10):1489–1495. doi: 10.1002/acr.22360 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Kim SC, Jin Y, Lee YC, Lii J, Franklin PD, Solomon DH, Franklin JM, Katz JN, Desai RJ. Association of Preoperative Opioid Use With Mortality and Short-term Safety Outcomes After Total Knee Replacement. JAMA Netw Open. 2019;2(7):e198061. doi: 10.1001/jamanetworkopen.2019.8061 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Chou R, Fanciullo GJ, Fine PG, Adler JA, Ballantyne JC, Davies P, Donovan MI, Fishbain DA, Foley KM, Fudin J, Gilson AM, Kelter A, Mauskop A, O’Connor PG, Passik SD, Pasternak GW, Portenoy RK, Rich BA, Roberts RG, Todd KH, Miaskowski C, American Pain Society-American Academy of Pain Medicine Opioids Guidelines Panel. Clinical guidelines for the use of chronic opioid therapy in chronic noncancer pain. J Pain Off J Am Pain Soc. 2009;10(2):113–130. doi: 10.1016/j.jpain.2008.10.008 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Goesling J, Moser SE, Zaidi B, Hassett AL, Hilliard P, Hallstrom B, Clauw DJ, Brummett CM. Trends and predictors of opioid use after total knee and total hip arthroplasty. Pain. 2016;157(6):1259–1265. doi: 10.1097/j.pain.0000000000000516 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Lauffenburger JC, Franklin JM, Krumme AA, Shrank WH, Brennan TA, Matlin OS, Spettell CM, Brill G, Choudhry NK. Longitudinal Patterns of Spending Enhance the Ability to Predict Costly Patients: A Novel Approach to Identify Patients for Cost Containment. Med Care. 2017;55(1):64–73. doi: 10.1097/MLR.0000000000000623 [DOI] [PubMed] [Google Scholar]
- 11.Singh JA, Lemay CA, Nobel L, Yang W, Weissman N, Saag KG, Allison J, Franklin PD. Association of Early Postoperative Pain Trajectories With Longer-term Pain Outcome After Primary Total Knee Arthroplasty. JAMA Netw Open. 2019;2(11):e1915105. doi: 10.1001/jamanetworkopen.2019.15105 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Franklin JM, Gopalakrishnan C, Krumme AA, Singh K, Rogers JR, Kimura J, McKay C, McElwee NE, Choudhry NK. The relative benefits of claims and electronic health record data for predicting medication adherence trajectory. Am Heart J. 2018;197:153–162. doi: 10.1016/j.ahj.2017.09.019 [DOI] [PubMed] [Google Scholar]
- 13.Franklin JM, Krumme AA, Tong AY, Shrank WH, Matlin OS, Brennan TA, Choudhry NK. Association between trajectories of statin adherence and subsequent cardiovascular events. Pharmacoepidemiol Drug Saf. 2015;24(10):1105–1113. doi: 10.1002/pds.3787 [DOI] [PubMed] [Google Scholar]
- 14.Franklin JM, Shrank WH, Pakes J, Sanfélix-Gimeno G, Matlin OS, Brennan TA, Choudhry NK. Group-based trajectory models: a new approach to classifying and predicting long-term medication adherence. Med Care. 2013;51(9):789–796. doi: 10.1097/MLR.0b013e3182984c1f [DOI] [PubMed] [Google Scholar]
- 15.Karhade AV, Schwab JH, Bedair HS. Development of Machine Learning Algorithms for Prediction of Sustained Postoperative Opioid Prescriptions After Total Hip Arthroplasty. J Arthroplasty. 2019;34(10):2272–2277.e1. doi: 10.1016/j.arth.2019.06.013 [DOI] [PubMed] [Google Scholar]
- 16.Politzer CS, Kildow BJ, Goltz DE, Green CL, Bolognesi MP, Seyler TM. Trends in Opioid Utilization Before and After Total Knee Arthroplasty. J Arthroplasty. 2018;33(7S):S147-S153.e1. doi: 10.1016/j.arth.2017.10.060 [DOI] [PubMed] [Google Scholar]
- 17.Data Resources | Drug Overdose | CDC Injury Center. Published October 9, 2020. Accessed November 17, 2020. https://www.cdc.gov/drugoverdose/resources/data.html
- 18.Friedman J, Hastie T, Tibshirani R. glmnet: Lasso and elastic-net regularized generalized linear models. R Package Version. 2009;1. [Google Scholar]
- 19.Gagne JJ, Glynn RJ, Avorn J, Levin R, Schneeweiss S. A combined comorbidity score predicted mortality in elderly patients better than existing scores. J Clin Epidemiol. 2011;64(7):749–759. doi: 10.1016/j.jclinepi.2010.10.004 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Kim DH, Glynn RJ, Avorn J, Lipsitz LA, Rockwood K, Pawar A, Schneeweiss S. Validation of a Claims-Based Frailty Index Against Physical Performance and Adverse Health Outcomes in the Health and Retirement Study. J Gerontol Ser A. 2019;74(8):1271–1276. doi: 10.1093/gerona/gly197 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Davis J, Goadrich M. The relationship between Precision-Recall and ROC curves. In: Proceedings of the 23rd International Conference on Machine Learning. ICML ‘06. Association for Computing Machinery; 2006:233–240. doi: 10.1145/1143844.1143874 [DOI] [Google Scholar]
- 22.Steyerberg EW. HARRELL FRANKE Jr., Regression Modeling Strategies: With Applications, to Linear Models, Logistic and Ordinal Regression, and Survival Analysis, 2nd ed. Heidelberg: Springer.: Book Review. Biometrics. 2016;72(3):1006–1007. doi: 10.1111/biom.12569 [DOI] [Google Scholar]
- 23.Steyerberg EW, Vickers AJ, Cook NR, Gerds T, Gonen M, Obuchowski N, Pencina MJ, Kattan MW. Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiol Camb Mass. 2010;21(1):128–138. doi: 10.1097/EDE.0b013e3181c30fb2 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Desai RJ, Wang SV, Vaduganathan M, Evers T, Schneeweiss S. Comparison of Machine Learning Methods With Traditional Models for Use of Administrative Claims With Electronic Medical Records to Predict Heart Failure Outcomes. JAMA Netw Open. 2020;3(1):e1918962-e1918962. doi: 10.1001/jamanetworkopen.2019.18962 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Karhade AV, Cha TD, Fogel HA, Hershman SH, Tobert DG, Schoenfeld AJ, Bono CM, Schwab JH. Predicting prolonged opioid prescriptions in opioid-naïve lumbar spine surgery patients. Spine J Off J North Am Spine Soc. Published online December 31, 2019. doi: 10.1016/j.spinee.2019.12.019 [DOI] [PubMed] [Google Scholar]
- 26.Lippold KM, Jones CM, Olsen EO, Giroir BP. Racial/Ethnic and Age Group Differences in Opioid and Synthetic Opioid-Involved Overdose Deaths Among Adults Aged ≥18 Years in Metropolitan Areas - United States, 2015–2017. MMWR Morb Mortal Wkly Rep. 2019;68(43):967–973. doi: 10.15585/mmwr.mm6843a3 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Shiels MS, Freedman ND, Thomas D, Berrington de Gonzalez A. Trends in U.S. Drug Overdose Deaths in Non-Hispanic Black, Hispanic, and Non-Hispanic White Persons, 2000–2015. Ann Intern Med. 2018;168(6):453–455. doi: 10.7326/M17-1812 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Burke LG, Zhou X, Boyle KL, Orav EJ, Bernson D, Hood M-E, Land T, Bharel M, Frakt AB. Trends in opioid use disorder and overdose among opioid-naive individuals receiving an opioid prescription in Massachusetts from 2011 to 2014. Addict Abingdon Engl. Published online November 6, 2019. doi: 10.1111/add.14867 [DOI] [PubMed] [Google Scholar]
- 29.Kim DH, Lin Y, Goytizolo EA, Kahn RL, Maalouf DB, Manohar A, Patt ML, Goon AK, Lee Y-Y, Ma Y, Yadeau JT. Adductor canal block versus femoral nerve block for total knee arthroplasty: a prospective, randomized, controlled trial. Anesthesiology. 2014;120(3):540–550. doi: 10.1097/ALN.0000000000000119 [DOI] [PubMed] [Google Scholar]
- 30.Krishnan SH, Gilbert LA, Ghoddoussi F, Applefield DJ, Kassab SS, Ellis TA. Addition of buprenorphine to local anesthetic in adductor canal blocks after total knee arthroplasty improves postoperative pain relief: a randomized controlled trial. J Clin Anesth. 2016;33:432–437. doi: 10.1016/j.jclinane.2016.04.021 [DOI] [PubMed] [Google Scholar]
- 31.Hah J, Mackey SC, Schmidt P, McCue R, Humphreys K, Trafton J, Efron B, Clay D, Sharifzadeh Y, Ruchelli G, Goodman S, Huddleston J, Maloney WJ, Dirbas FM, Shrager J, Costouros JG, Curtin C, Carroll I. Effect of Perioperative Gabapentin on Postoperative Pain Resolution and Opioid Cessation in a Mixed Surgical Cohort: A Randomized Clinical Trial. JAMA Surg. 2018;153(4):303–311. doi: 10.1001/jamasurg.2017.4915 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Derogatis MJ, Sodhi N, Anis HK, Ehiorobo JO, Bhave A, Mont MA. Pain Management Strategies To Reduce Opioid Use Following Total Knee Arthroplasty. Surg Technol Int. 2019;35:301–310. [PubMed] [Google Scholar]
- 33.West SL, Savitz DA, Koch G, Strom BL, Guess HA, Hartzema A. Recall accuracy for prescription medications: self-report compared with database information. Am J Epidemiol. 1995;142(10):1103–1112. doi: 10.1093/oxfordjournals.aje.a117563 [DOI] [PubMed] [Google Scholar]
- 34.West SL, Strom BL, Freundlich B, Normand E, Koch G, Savitz DA. Completeness of prescription recording in outpatient medical records from a health maintenance organization. J Clin Epidemiol. 1994;47(2):165–171. doi: 10.1016/0895-4356(94)90021-3 [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.