Frailty Indicator Using a Well- Established Frailty Phenotype Jodi - PowerPoint PPT Presentation

Development of a Claims-Based Frailty Indicator Using a Well- Established Frailty Phenotype Jodi B. Segal, MD, MPH, Hsien Yen-Chang, PhD, Yu Du, MS, Michelle Carson, PhD, Jeremy Walston, MD, Ravi Varadhan, PhD Funding: National Institute on Aging, R21 AG048494-01

Background • In 2001, Fried et al described a frailty phenotype in the Cardiovascular Health Study (CHS) • The phenotype is manifest when 3 or more of the following are present: – Low grip strength – Low energy – Slowed walking speed – Low physical activity – Unintentional weight loss (>=10 lb) • Simple measures, but not routinely measured or recorded in most clinical encounters

Goals • To develop a Claims-based Frailty Indicator that will identify people with frailty using only administrative claims data • To validate this measure of frailty as a predictor of clinical outcomes • To compare performance of this measure as a predictor of clinical outcomes to the original frailty phenotype measure

Methods • Data : Cardiovascular Health Study (CHS) cohort previously linked to Medicare claims – 5,201 men and women from 4 U.S. communities (starting in 1989) – 687 more African Americans joined later – Examined annually through 1999, with phone calls every six months after 1999 – Medicare data from 1992-2013 linked to CHS data • Frailty reference standard – >3 indicators = Frail – Fewer than 3 = Non-frail

Methods Candidate variable selection – Used literature to identify variables previously used to classify individuals as frail or disabled with claims data or EMR – Added other variables from AHRQ’s Clinical Classifications Software (CCS) using ICD-9-CM codes – Operationalized them to be identifiable in claims data

Methods Used claims from 6 months before the 5 th and 9 th year study visits (individuals contributed twice) claims claims Visit 5 (9 th yr) Visit 2 (5 th yr) Measured frailty Measured frailty

Methods Modeling frailty – Adaptive Lasso technique with logistic regression – Optimal Lasso penalty chosen via 10-fold cross- validation to maximize area under the receiver operator curve – Also tested a gradient boosting method, a random forest method, and logic (Boolean) regression – We required continuous enrollment in Medicare Parts A & B during the 6 month windows of interest

Methods Internal Validation – Operationalized outcomes of fracture, death, nursing home admission, hospitalization and disability occurrence in the 5 years after 5 th year study visit – Modeled the risk of events for frail and non-frail classified at a cutoff that had good specificity and acceptable sensitivity, also as a continuous variable claims data claims cohort data Visit 2 5 year interval

Results Characteristics of Participants in CHS with Continuous Enrollment (n=4454) VARIABLE % Age 72 years (mean) White 84.1% Black 15.4% Other 0.46% Female 58.8% Married 66.2% Widowed 24.6% Divorced 3.89% Separated 0.90% Never Married 4.47%

Results B- coefficient Variable 1.24 Impaired mobility 0.54 Depression 0.50 Congestive Heart Failure Variables Parkinson’s disease 0.50 Retained in -0.49 White race 0.43 Arthritis (any type) Claims-Based 0.33 Cognitive impairment 0.31 Charlson comorbidity index (>0, 0) Frailty 0.28 Stroke 0.24 Paranoia Indicator 0.23 Chronic skin ulcer 0.21 Pneumonia -0.19 Male sex 0.18 Skin and soft tissue infection 0.14 Mycoses 0.09 Age (in 5 year categories) 0.09 Admission in past 6 months 0.08 Gout or other crystal-induced arthropathy 0.08 Falls 0.05 Musculoskeletal problems 0.05 Urinary tract infection

Results At cutoff of 0.12, Sens=0.66 Spec=0.73 PPV =0.24 NPV=0.94 At cutoff of 0.3, Sens =0.16 Spec=0.98 PPV = 0.46 NPV = 0.90 For every 10% absolute Area under increase in predicted ROC=0.75 probability of frailty, the odds of actual frailty increase by 2.4 - fold

Classification Actual frailty vs. Predicted frailty (p>=0.2) - Enrollment-Eligible Patients Actual 0 1 Total Predicted 0 3519 258 3912 1 324 156 542 Total 3843 414 4454

Results : Internal Validation Outcome Measure N Unadjusted [95% C.I.] Adjusted [95% C.I.]* 3.81 [3.15 – 4.62] 1.81 [1.41 – 2.31] Death within 5 years OR 4453 3.18 [2.72 – 3.71] 1.61 [1.30 – 2.00] Time to death HR 4453 2.18 [1.67 – 2.86] 1.46 [1.07 – 1.99] Admission within 5 years OR 2875 1.71 [1.46 – 1.99] 1.30 [1.06 – 1.58] Time to first admission HR 2875 0.97 [0.76 – 1.25] Fracture within 5 years** OR 4255 1.18 [0.95 - 1.45] 1.45 [1.04 – 2.01] Nursing home admission OR 4209 3.80 [2.96 - 4.88] 3.56 [2.90 – 4.37] 2.15 [1.69 – 2.74] Disability within 5 years OR 4184 *adjusted for sex and age (in years), **counting first fracture per year of each eligible body part,***two stage model – estimate reported is impact of frailty among people with any fractures, OR=odds ratio, HR=hazard ratio

For Every 10% Increase in Predicted Probability of Frailty Outcome Measure N Unadjusted [95% C.I.] Adjusted [95% C.I.]* 1.86 [1.71 – 2.00] 1.49 [1.33 – 1.68] Death within 5 years OR 4453 1.55 [1.47 – 1.63] 1.34 [1.23 – 1.45] Time to death HR 4453 1.64 [1.55 – 2.04] 1.93 [1.60 – 2.32] Admission within 5 years OR 2875 1.42 [1.36 – 1.51] 1.52 [1.41 – 1.64] Time to first admission HR 2875 1.08 [1.00 – 1.17] 0.96 [0.86 – 1.08] Fracture within 5 years** OR 4255 1.95 [1.69 – 2.02] 1.38 [1.21 – 1.58] Nursing home admission OR 4209 2.06 [1.87 – 2.26] 2.04 [1.77 – 2.36] Disability within 5 years OR 4184 *adjusted for sex and age (in years), **counting first fracture per year of each eligible body part,***two stage model – estimate reported is impact of frailty among people with any fractures, OR=odds ratio, HR=hazard ratio

Comparing Age (5 years)* and Predicted Frailty** (10% increase) Death with 5 years (OR) Time to death (HR) Admission within 5 years (OR) Time to first admission (HR) Fracture within 5 years (OR) Age by 5 years Nursing home admission (OR) Predicted Frailty by 10% Disability within 5 years (OR) 0.1 1 10 Measure of Association *sex and frailty adjusted, **sex and age adjusted

Comparison to Measured Frailty (Phenotype) Outcome Measure N Claims-based Measured Frailty Indicator Phenotype Unadjusted [95% Unadjusted C.I.]* [95% C.I.]* 3.81 [3.15 – 4.62] 3.26 [2.62 – 4.05] Death within 5 years OR 4453 3.18 [2.72 – 3.71] 2.82 [2.36 – 3.37] Time to death HR 4453 2.18 [1.67 – 2.86] 2.76 [2.02 – 3.77] Admission within 5 years OR 2875 1.71 [1.46 – 1.99] 1.95 [1.65 – 2.30] Time to first admission HR 2875 1.18 [0.95 – 1.45] 1.21 [0.96 – 1.53] Fracture within 5 years** OR 4255 3.80 [2.96 – 4.88] 3.35 [2.52 – 4.45] Nursing home admission OR 4209 3.56 [2.90 – 4.37] 4.30 [3.40 – 5.45] Disability within 5 years OR 4184

Comparison to Measured Frailty (Phenotype) Outcome Measure N Claims-based Measured Frailty Indicator Phenotype Adjusted [95% C.I.]* Adjusted [95% C.I.]* 1.81 [1.41 – 2.31] 2.46 [1.94 – 3.11] Death within 5 years OR 4453 1.61 [1.30 – 2.00] 2.07 [1.70 – 2.53] Time to death HR 4453 1.46 [1.07 – 1.99] Admission within 5 years OR 2875 2.46 [1.79 - 3.39] 1.30 [1.06 – 1.58] 1.76 [1.47 – 2.10] Time to first admission HR 2875 0.97 [0.76 – 1.25] 1.09 [0.85 – 1.38] Fracture within 5 years** OR 4255 1.45 [1.04 – 2.01] Nursing home admission OR 4209 2.08 [1.53 - 2.83] 2.15 [1.69 – 2.74] 3.39 [2.66 – 4.32] Disability within 5 years OR 4184 *adjusted for sex and age (in years), **counting first fracture per year of each eligible body part,***two stage model – estimate reported is impact of frailty among people with any fractures, OR=odds ratio, HR=hazard ratio, IRR=incidence rate ratio

Comparing CFI and Phenotype, Unadjusted Death with 5 years (OR) Time to death (HR) Admission within 5 years (OR) Time to first admission (HR) Frailty Phenotype Claims based Frailty Indicator Fracture within 5 years (OR) Nursing home admission (OR) Disability within 5 years (OR) 0.1 1 10 Measure of Association

Comparing CFI and Phenotype, Age and Sex Adjusted Death with 5 years (OR) Time to death (HR) Admission within 5 years (OR) Time to first admission (HR) Frailty Phenotype Claims based Frailty Indicator Fracture within 5 years (OR) Nursing home admission (OR) Disability within 5 years (OR) 0.1 1 10 Measure of Association

Additional Results • Random forest method with slightly higher area under ROC curve, at the expense of transparency • Higher area under ROC curve than commonly used Frailty Index (count of comorbidities) • Charlson Comorbidity Index alone is nicely predictive of some of the tested outcomes

Conclusions • Adaptive lasso technique produced a model that identifies individuals as frail using claims data alone • Claims-based Frailty Index predicts outcomes within same data set similarly to the frailty phenotype

Limitations • Not yet validated in an external data set • Perhaps there are other covariates that would further improve classification

Implications • Claims-only index built against the accepted reference standard for measured frailty has good predictive value • Claims-based Indicator of Frailty has broad uses for: – Risk adjustment – Exploration of heterogeneity of treatment effect – Population health management – Emergency preparedness

THANK YOU