Скачать презентацию Use of Propensity Score Methods in Health Care Скачать презентацию Use of Propensity Score Methods in Health Care

d2c0e05079141aa0ba607984097c2538.ppt

  • Количество слайдов: 20

Use of Propensity Score Methods in Health Care Data Networks: Current Practice and Challenges Use of Propensity Score Methods in Health Care Data Networks: Current Practice and Challenges Jennifer Clark Nelson, Ph. D Biostatistics Unit, Kaiser Permanente Washington Health Research Institute (KPWHRI) Department of Biostatistics, University of Washington Kaiser. International Conference on Health Policy Permanente Washington Health Research Institute Statistics, January 12, 2018

Overview • Big (Health Care) Data 101 • National safety surveillance initiatives in the Overview • Big (Health Care) Data 101 • National safety surveillance initiatives in the U. S. o Vaccine Safety Datalink (CDC) o Sentinel Initiative (FDA) • Two example safety evaluations using propensity score methods • Challenges 19 March 2018

Big (Health Care) Data 101 § What is it? o Data collected by public Big (Health Care) Data 101 § What is it? o Data collected by public and private organizations for the purposes of registration, transaction and record keeping, usually during the delivery of health care services o Also called administrative, clinical, or electronic health care data § How does it generated? o Health care system encounters (outpatient, inpatient, pharmacy) create electronic claims to the payer for reimbursement o Paper or electronic health record (EHR) captures standard medical and clinical data gathered in one provider’s office § What kind of information is collected? o Diagnosis codes (ICD-9), procedure codes (CPT/HCPCS), dates o Pharmacy dispensings (drug name, dates, etc. ) o Patient demographics, care setting, type of provider… 19 March 2018

Big (Health Care) Data 101 § How can it be re-purposed for research? o Big (Health Care) Data 101 § How can it be re-purposed for research? o Link data across various sources (e. g. , clinical records, billing data, lab results, insurance claims, and vitals) o Commonly define these data elements (common data model) o Link them across data partners for multi-site research o An integrated picture of health/healthcare emerges for large cohorts § Is this a new idea? o No (HMORN -- now Health Care Systems Research Network, Medicare data have been used for decades for research) but… o Their use is rapidly expanding due to development of new shared national multi-purpose big data networks § Is this a good idea? 4 o Yes (we need efficient and cost-effective ways to fill evidence 19 March 2018 gaps left by traditional RCTs and observational studies) but it

Vaccine Safety Datalink (VSD) § Established in 1990 § A collaborative project among CDC Vaccine Safety Datalink (VSD) § Established in 1990 § A collaborative project among CDC and 8 integrated health care delivery systems § Priorities – To evaluate the safety of newly licensed vaccines or new recommendations for existing vaccines – To develop & evaluate new methods for safety § Pre-2004: traditional retrospective studies (~12 month lag) § Post-2004: near-real time evaluations (weekly updates)

Kaiser Permanente Washington Northwest Kaiser Permanente VSD Sites Health Partners Marshfield Clinic No. CA Kaiser Permanente Washington Northwest Kaiser Permanente VSD Sites Health Partners Marshfield Clinic No. CA Kaiser Permanente Harvard Pilgrim Kaiser Permanente Colorado So. CA Kaiser Permanente §Captures medical care and vaccination data §Large population: ~10 million ( or 3% of U. S. population) §Population-based: members from 7 health plans

VSD Data Sources and Elements Hospital discharge diagnosis codes Enrollment and demographics ER Visits VSD Data Sources and Elements Hospital discharge diagnosis codes Enrollment and demographics ER Visits Procedure Codes Linked by Study IDs Birth and death certificate information & Family Linkage Outpatient and Clinic visits Immunizations Records

Vision for Sentinel “…a national electronic system that will transform FDA’s ability to track Vision for Sentinel “…a national electronic system that will transform FDA’s ability to track the safety of drugs, biologics, and medical devices once they reach the market. ” “…aims to develop and implement a proactive system that will complement existing systems that the Agency has in place to track reports of adverse events. ” “…enables FDA to actively query diverse automated healthcare data holders—like EHR systems, administrative and insurance claims databases, and registries—to evaluate possible medical product safety issues quickly and securely. ” http: //www. fda. gov/Safety/FDAs. Sentinel. Initiative 8

All Partners Bring Expertise, Data Partners Respond to Queries Lead – HPHC Institute Data All Partners Bring Expertise, Data Partners Respond to Queries Lead – HPHC Institute Data and scientific partners Scientific partners Institute for Health [email protected] org © 2017 Sentinel Operations Center. All Rights Reserved. 9

Sentinel is a Distributed Data Network info@sentinelsystem. org © 2017 Sentinel Operations Center. All Sentinel is a Distributed Data Network [email protected] org © 2017 Sentinel Operations Center. All Rights Reserved. 10

Prospective Surveillance Pilot of Rivaroxaban Safety § Leads: E Chrischilles, R Carnahan (U of Prospective Surveillance Pilot of Rivaroxaban Safety § Leads: E Chrischilles, R Carnahan (U of Iowa) § Co-investigators: J Gagne, B Fireman, J Nelson, S Toh, A Shoaibi, M Reichman, S Wang, M Nguyen, R Zhang, R Izem, M Goulding, MR Southworth, D Graham, C Fuller, H Katcoff, T Woodworth, C Rogers, R Saliga, N Lin, C Mc. Mahill-Walraven, V Nair, N Selvam § Conducted at 4 Sentinel Data Partners § Funding: This project was funded by the FDA through HHS Mini-Sentinel contract #HHSF 223200910006 I [email protected] org © 2017 Sentinel Operations Center. All Rights Reserved. 11

Study goals and design § Assess the safety of rivaroxaban (Xarelto®) vs. warfarin among Study goals and design § Assess the safety of rivaroxaban (Xarelto®) vs. warfarin among adults with atrial fibrillation in the drug’s early uptake period (beginning in Nov 2011) § Targeted outcomes of interest • GI bleed, ischemic stroke, intracranial hemorrhage § New user cohort followed • From day after first dispensing • Until outcome occurrence, treatment discontinuation, initiation of another anti-coagulant, death, disenrollment from health plan, or end of study data (whichever is first) [email protected] org © 2017 Sentinel Operations Center. All Rights Reserved. 12

Propensity Score (PS) Approach § § Variable ratio PS matching (up to 10 warfarin Propensity Score (PS) Approach § § Variable ratio PS matching (up to 10 warfarin users) Nearest neighbor algorithm, matching caliper 0. 05 PS estimation and matching within Data Partner Many (70+) confounders • • Age, sex, year of index date, combined comorbidity score Health service utilization (counts of encounters, # of drugs) Procedures and diagnoses: risk factors for bleeding, stroke Medications: oral cardiovascular agents, medications that increase bleeding risk, interacting medications § Cox regression stratified by Data Partner and matched set to estimate a hazard ratio [email protected] org © 2017 Sentinel Operations Center. All Rights Reserved. 13

Cohort Characteristics Selected Characteristics Gender (F) 41, 800 16, 374 (39. 2) 87, 907 Cohort Characteristics Selected Characteristics Gender (F) 41, 800 16, 374 (39. 2) 87, 907 37, 017 (42. 1) 0. 06 36, 173 14, 669 (40. 6) Matcheda N(%) Warfarin Standardized Difference 79, 520 14, 574 (40. 3) 0. 005 Age-mean (SD) 69. 7 (10. 7) 73. 4 (10. 6) 0. 352 71. 1 (10. 4) 71. 1 (10. 7) 0 Combined Comorbidity Score - mean (SD) Atrial fibrillation 2. 4 (2. 4) 3. 2 (2. 8) 0. 313 2. 5 (2. 4) 0. 007 36, 581 (87. 5) 77, 568 (88. 2) 0. 022 31, 630 (87. 4) 31, 866 (88. 1) 0. 02 Atrial flutter 7, 627 (18. 2) 12, 454 (14. 2) 0. 111 5, 994 (16. 6) 6, 008 (16. 6) 0. 001 GI bleed 1, 507 (3. 6) 4, 841 (5. 5) 0. 091 1, 393 (3. 9) 1, 426 (3. 9) 0. 005 Intracranial hemorrhage 231 (0. 6) 1, 152 (1. 3) 0. 079 224 (0. 6) 239 (0. 7) 0. 005 Ischemic stroke 3, 150 (7. 5) 10, 207 (11. 6) 0. 139 3, 031 (8. 4) 2, 973 (8. 2) 0. 006 Hypertension 32, 865 (78. 6) 71, 386 (81. 2) 0. 064 28, 662 (79. 2) 28, 683 (79. 3) 0. 001 Hyperlipidemia 12, 819 (30. 7) 25, 265 (28. 7) 0. 042 10, 884 (30. 1) 11, 046 (30. 5) 0. 01 Heart failure or 15, 110 (36. 1) cardiomyopathy Peripheral vascular disease 6, 638 (15. 9) 39, 359 (44. 8) 0. 176 13, 781 (38. 1) 13, 940 (38. 5) 0. 009 18, 645 (21. 2) 0. 137 6, 234 (17. 2) 6, 277 (17. 4) 0. 003 Diabetes 12, 505 (29. 9) 31, 905 (36. 3) 0. 136 11, 417 (31. 6) 11, 398 (31. 5) 0. 001 Venous thromboembolism Walker use 2, 525 (6. 0) 10, 598 (12. 1) 0. 211 2, 456 (6. 8) 2, 340 (6. 5) 0. 013 886 (2. 1) 3, 126 (3. 6) 0. 087 844 (2. 3) 807 (2. 2) 0. 007 Home oxygen 2, 240 (5. 4) 7, 017 (8. 0) 0. 105 2, 123 (5. 9) 2, 078 (5. 7) 0. 005 [email protected] org a weighted Unmatched N(%) Rivaroxaban N(%) Warfarin Standardized Difference N(%) Rivaroxaban © 2017 Sentinel Operations Center. All Rights Reserved. 14

End-of-Surveillance Results (Specific Criteria for Outcomes A) Person. Outcome/ Years at Comparator New Users End-of-Surveillance Results (Specific Criteria for Outcomes A) Person. Outcome/ Years at Comparator New Users Risk Gastrointestinal Bleeding Rivaroxaban 36, 173 8, 427 Warfarin 79, 520 15, 384 Ischemic Stroke Rivaroxaban 36, 512 8, 572 Warfarin 80, 180 15, 672 Intracranial Hemorrhage Rivaroxaban 36, 171 8, 502 Warfarin 79, 529 15, 551 Events 423 651 82 268 46 143 Adjusted Incidence Rate per 1000 Adjusted Hazard Person-Years. C Ratio (95% CI)B 1. 47 (1. 29, 1. 67) 50. 20 34. 82 9. 57 0. 61 (0. 47, 0. 79) 17. 10 5. 41 0. 71 (0. 50, 1. 01) 7. 49 A Outcome events required diagnosis codes in primary position. Monitoring period started Nov 1, 2011 for all data partners, but the end date varied among Data Partners: April 30, 2014, Dec 31, 2014, March 31, 2015, and April 30, 2015. Matching caliper for this analysis was 0. 01. B Hazard Ratios estimated by stratified Cox regression conditioned on Data Partner and PS matched set. Confidence intervals are nominal 95% intervals for the final hazard ratio estimates. C Incidence rates adjusted for censoring in matched sets and variable ratio matching. [email protected] org © 2017 Sentinel Operations Center. All Rights Reserved. 15

Pneumococcal Conjugate Vaccine Safety § Leads: HF Tseng, S Jacobsen (Kaiser Permanente So Cal) Pneumococcal Conjugate Vaccine Safety § Leads: HF Tseng, S Jacobsen (Kaiser Permanente So Cal) § Co-investigators: L Sy, L Qian, A Liu, C Mercado, B Lewin, S Tartof, JC Nelson, L Jackson, M Daley, E Weintraub, N Klein, E Belongia, EG Lilies § Conducted at 6 VSD Data Partners § Funding: This study was funded through the Vaccine Safety Datalink (VSD) under contract 200 -2012 -53580 from the Centers for Disease Control and Prevention

Study goals, design and PS approach § Examine a large cohort of adults 65+ Study goals, design and PS approach § Examine a large cohort of adults 65+ years for risk of adverse events (AEs) following vaccination with PCV 13 as compared to PPSV 23, a long-standing vaccine with a satisfactory safety profile § Targeted outcomes of interest – CVD events, Bell’s palsy, Guillan Barre syndrome, syncope, erythema, multiforme, thrombocytopenia, cellulitis & infection, allergic reaction, and anaphylaxis § Retrospective cohort of PCV 13 and PPSV 23 vaccinees, 2011 -2015 – Followed in post-vaccination risk windows (definitions varied by outcome) with censoring at disenrollment, death, or receipt of another vaccine § Confounders: age, sex, site, comorbidity score, health care utilization

Challenges 1. Lack of a controlled research setting § What PS methods can best Challenges 1. Lack of a controlled research setting § What PS methods can best handle a large # of confounders? 2. Safety outcomes are rare (e. g. , 1 per 10, 000 doses) § How do typical PS methods perform in small sample conditions? 3. Exposure may also be uncommon (early/limited new drug uptake) § How does this impact our ability/choice in method to estimate the PS? 4. Individual level data cannot typically be pooled (to protect privacy) § What PS methods can be adapted simply to deal with this constraint? 18 19 March 2018

References 1. Baggs J, Gee J, Lewis E et al. , The Vaccine Safety References 1. Baggs J, Gee J, Lewis E et al. , The Vaccine Safety Datalink: a model for monitoring immunization safety. Pediatrics. 2011; 127 Suppl 1: S 45 -53. 2. Behrman RE, Benner JS, Brown JS et al. , Developing the Sentinel System – a national resource for evidence development. N Engl J Med 2011; 364(6): 498 -9. 3. Chrischilles E, Gagne J, Fireman B et al. , Prospective pilot of rivaroxaban safety within the US Food and Drug Administration Sentinel Program (in press, Pharmacoepi Drug Saf) 4. Nelson JC, Cook AJ, Yu O, et al. Challenges in the design and analysis of sequentially monitored postmarket safety surveillance evaluations using electronic observational health care data. Pharmacoepidemiology and drug safety 2012; 21 Suppl 1: 62 -71. 5. Routine Querying System. 2016. at https: //www. sentinelinitiative. org/sites/default/files/Surveillance. Tool s/Routine. Querying/Sentinel-Routine_Querying_System-

Contact Information Jennifer Clark Nelson Biostatistics Unit Kaiser Permanente Washington Health Research Institute nelson. Contact Information Jennifer Clark Nelson Biostatistics Unit Kaiser Permanente Washington Health Research Institute nelson. [email protected] org 20 19 March 2018