A A Mach chin ine Le Learnin ing Alg lgorit ithm for Sh Short-Term Mortali lity Predic ictio ion in in Cancer Patie ient Popula latio ions Jun June 25, 2017 Maximilian J. Pany MD-PhD candidate, Harvard/MIT Harvard Medical School Aymen A. Elfiky Ravi B. Parikh Ziad Obermeyer
Outline 1. Background & Objective 2. Methods 3. Results 4. Conclusions
Outline 1. Background & Objective 2. Methods 3. Results 4. Conclusions
Chemotherapy is life-saving, but not costless • Chemotherapy is started too often, too late - Patients who die soon after starting chemotherapy incur the costs of treatment without its benefits • Understanding the risks of chemotherapy is important - Informed consent to treatment - Finances, estate planning - Family, life events - Palliative care, advance care directives • Chemotherapy use near the end of life as a marker of poor quality of care
Our prognoses are only as accurate as our predictions , and predictions are hard • Side effects of chemotherapy are variable - Genetics, comorbidities • Cognitive biases lead to underestimated mortality risk - Optimism, avoidance, incentives, … • Currently, mortality estimates rely on RCT and SEER data https://clinicaltrials.gov/; https://seer.cancer.gov
Can a machine learning model accurately predict mortality at chemotherapy initiation? • Fundamental challenge and opportunity - Rich clinical data from EHR - Structured and unstructured data elements - All patients in given clinical setting • Objective : To develop and validate a ML algorithm predicting mortality for patients initiating chemotherapy regimens at a national cancer center - Critical event in the disease trajectory - ‘Pause point’ to weigh difficult questions
Outline 1. Background & Objective 2. Methods 3. Results 4. Conclusions
Empirical approach • Track all ~27,000 patients from a national cancer center at start of ~52,000 chemotherapy regimens • Predict mortality (Social Security data) at 30 days derivation validation 2004 2012 2014 • Compare to https://pinterest.com; www.happyfamilyart.com
RHS: 5,390 predictors chosen by cross-validation • Categories Demographics Vital signs Care utilization Laboratory results Grouped ICD-9 codes Natural language processing (diagnoses, procedures) of physician notes Procedures Prescribed medications • Two time periods before chemo start - 0-1 month (recent) - 1-12 months (baseline) • Data up to day before included
Outline 1. Background & Objective 2. Methods 3. Results 4. Conclusions
Model predictors Mean by risk decile Model Variable Type Top Median Bottom variance explained Demographics Age Mean 62.3 62.1 51.9 1.30% Female Mean 0.56 0.61 0.87 1.17% Diagnoses Linear terms: 14% Comorbidity score Year, max 4.15 2.62 1.59 0.01% Non-linear terms and Ascites Year, max 0.31 0.07 0.01 0.39% Medications interactions: 86% Corticosteroids Year, max 0.53 0 0 0.15% Vital signs Pulse Year, max 106.1 95.7 87.1 0.37% Weight (kg) Change -3.1 -1 0.1 0.00% Labs WBC Year, max 13.9 12.4 9.8 0.03% C-reactive protein Year, max 93.9 65.6 2.2 0.19% Diagnostic testing Ejection fraction (%) Year, max 54.4 48 51.9 0.01%
Survival (validation set) by predicted risk q1 100 q2 Palliative chemo 75 30-day mortality (AUC 0.94) mean Observed survival (%) • Highest risk decile: 22.6% • Lowest risk decile: 0.0% q9 50 Accurate for • All cancers, all stages (AUC>0.90 for all) 25 q10 • New DFCI trial agents not in derivation set (AUC 0.94) 0 0d 30d 90d 180d
ML predictions vs RCT mean mortality by regimen Observed mortality (%) RCT ML Regimen- specific mortality Predicted mortality (%)
Predictions vs SEER estimates Observed mortality (%) SEER ML Stage 4; Age, sex, race-specific estimates Predicted mortality (%)
Outline 1. Background & Objective 2. Methods 3. Results 4. Conclusions
Conclusions and next steps • A machine learning model accurately predicted short- term mortality in patients initiating chemotherapy • The model performed better than commonly used trial- and population-based estimates - Implications for care and financial planning • Further research, including prospective studies, is necessary to determine this model’s generalizability
Limitations • Single-institution study based on retrospective data • Applicability to novel oncologic therapies? • Model trained only on patients selected into chemotherapy • Counter-factual: we don’t know what would have happened without chemotherapy
Acknowledgments • Co-authors: Drs. Aymen A. Elfiky, Ravi B. Parikh, and Ziad Obermeyer • Comments: Drs. Jennifer Temel and Deborah Schrag • Funding : Office of the Director of the National Institutes of Health (DP5 OD012161) National Institute of Aging (T32 AG51108) Dana Farber Cancer Institute
References (1/2) • Emanuel EJ, Young-Xu Y, Levinsky NG, Gazelle G, Saynina O, Ash AS. Chemotherapy use among Medicare beneficiaries at the end of life. Ann Intern Med 2003;138(8):639 – 43. • Earle CC, Neville BA, Landrum MB, Ayanian JZ, Block SD, Weeks JC. Trends in the Aggressiveness of Cancer Care Near the End of Life. J Clin Oncol 2004;22(2):315 – 21. • Earle CC, Landrum MB, Souza JM, Neville BA, Weeks JC, Ayanian JZ. Aggressiveness of Cancer Care Near the End of Life: Is It a Quality-of-Care Issue? J Clin Oncol 2008;26(23):3860 – 6. • Saito AM, Landrum MB, Neville BA, Ayanian JZ, Earle CC. The effect on survival of continuing chemotherapy to near death. BMC Palliat Care 2011;10:14. • Prigerson HG, Bao Y, Shah MA, et al. Chemotherapy Use, Performance Status, and Quality of Life at the End of Life. JAMA Oncol 2015;1(6):778 – 84. • Schnipper LE, Smith TJ, Raghavan D, et al. American Society of Clinical Oncology identifies five key opportunities to improve care and reduce costs: the top five list for oncology. J Clin Oncol Off J Am Soc Clin Oncol 2012;30(14):1715 – 24. • National Quality Forum. Cancer Measures [Internet]. Washington (DC): 2012. Available from: https://www.qualityforum.org/News_And_Resources/Endorsement_Summaries/ Cancer_Measures_Endorsement_Summary.aspx • Greer JA, Pirl WF, Jackson VA, et al. Effect of early palliative care on chemotherapy use and end-of-life care in patients with metastatic non-small-cell lung cancer. J Clin Oncol Off J Am Soc Clin Oncol 2012;30(4):394 – 400.
References (2/2) • Glare P, Virik K, Jones M, et al. A systematic review of physicians’ survival predictions in terminally ill cancer patients. BMJ 2003;327(7408):195. • Stone PC, Lund S. Predicting prognosis in patients with advanced cancer. Ann Oncol 2007;18(6):971 – 6. • Silvestri G, Pritchard R, Welch HG. Preferences for chemotherapy in patients with advanced non-small cell lung cancer: descriptive study based on scripted interviews. BMJ 1998;317(7161):771 – 5. • Hirose T, Yamaoka T, Ohnishi T, et al. Patient willingness to undergo chemotherapy and thoracic radiotherapy for locally advanced non-small cell lung cancer. Psychooncology 2009;18(5):483 – 9. • Keating NL, Landrum MB, Rogers SO, et al. Physician factors associated with discussions about end-of-life care. Cancer 2010;116(4):998 – 1006. • Keating NL, Beth Landrum M, Arora NK, et al. Cancer patients’ roles in treatment decisions: do characteristics of the decision influence roles? J Clin Oncol Off J Am Soc Clin Oncol 2010;28(28):4364 – 70. • Liu P- H, Landrum MB, Weeks JC, et al. Physicians’ propensity to discuss prognosis is associated with patients’ awareness of prognosis for metastatic cancers. J Palliat Med 2014;17(6):673 – 82. • Statistical Summaries - SEER Cancer Statistics [Internet]. Available from: https://seer.cancer.gov/statistics/summaries.html
Recommend
More recommend
