- Slides: 23
Machine Learning Basics: Advice and Resources for Clinical Researchers Michael P. Cary, Jr. , Ph. D, RN Associate Professor, Duke University School of Nursing Interdisciplinary Colloquium for Outcomes Research (ISCORe)
Agenda What is machine learning? Why should clinical researchers care? What were some challenges and opportunities to incorporating machine learning into my program of research?
What is Machine Learning? a mechanically, electrically, or electronically operated device designed to perform a task. learning is the activity or process of gaining knowledge by studying, practicing, being taught, or experiencing something. https: //www. sas. com/content/dam/SAS/en_us/doc/whitepaper 1/machine-learningprimer-108796. pdf
Machines Learning, Really? Machines learn by studying data to detect patterns or by applying known rules to: • Categorize like people or things • Predict outcomes based on identified patterns • Identify unknown patterns and relationships https: //www. sas. com/content/dam/SAS/en_us/doc/whitepaper 1/machine-learningprimer-108796. pdf
Why should clinicians care? BIG DATA = BIG OPPORTUNITIES
As a clinician, what does machine learning mean to me? Big Data ML Algorithms Clinical Insight/ Informed Decisions
Overview of Research Program Two interrelated areas: 1 Preventing functional decline and rehospitalization among vulnerable populations following post-acute rehabilitation. 2 Using quality measures to promote system-level improvements in postacute care (PAC) settings to achieve better patient outcomes.
Why Hip Fractures? Over 90% of older adults admitted to the hospital for a hip fracture use Post-Acute Care for rehabilitation and recovery 29% 70% 25% die within 1 year do not regain independent mobility are permanently admitted to nursing homes
Why Inpatient Rehabilitation? Inpatient Rehabilitation Facilities (IRFs) provide intensive rehabilitation services Little evidence about: • Specific individual characteristics of orthopedic patients (demographic and clinical) • Vulnerable subgroups (e. g. , racial/ethnic minorities)
Study 1: Individual Characteristics and PAC Outcomes Examine the influence of individual characteristics • Predisposing: Age, Race, Sex • Enabling: Social Support (live with others) • Need: Functional Status at Admission, Comorbidities --- ON --PAC outcomes (patient-level) • Functional Status at Discharge • Community Discharge (home)
Methods Design: Retrospective cohort study Setting: 1, 112 Medicare certified IRFs in the US Population: Adults 65 yrs. and older, Medicare fee-for-service insurance, with a fractured hip in 2009 Sample: Excluded if: (1) did not live at home prior to acute care hospital admission; (2) stayed in the IRF for <3 days or >30 days after admission; or (3) if they died during the rehabilitation stay. Final sample size was N=34, 984.
Analysis Multivariate regression analysis was used to determine the relationships between independent variables and functional status at discharge. Logistic regression analysis was used to determine which independent variables were associated with discharged home.
Findings PAC outcomes in this national study were predicted by a variety of individual characteristics, most notably enabling and need factors. Subgroups at risk for poorer outcomes: • • Older patients Minorities (Blacks, Hispanics) Men Persons with greater disease severity (increase number of comorbidities) Cary MP Jr, et al, (2016). Inpatient rehabilitation outcomes in a national sample of Medicare beneficiaries with hip fracture. Journal of Applied Gerontology, 35(1), 62 -83.
Roadblocks to Improving PAC Hip Fracture Outcomes Comorbidity measures included simple counts No real-time data
Next Steps: Hip Fractures, Comorbidities, and PAC Outcomes Poor health outcomes increase with the number of comorbid conditions Little guidance is offered to clinicians in adjusting care for IRF patients with MCCs Readmissions are high among IRF patients; most risk prediction models perform poorly
Study 2: MCCs and Hospital Readmission Aim 1 Aim 2 Identify clusters of MCCs associated with hospital readmission among hip fracture patients discharged from IRFs. Assess the performance of the traditional statistical model (Aim 1 a) and the machinelearning model (Aim 1 b) using receiver operating curves and the area under the curve as well as other measures of accuracy. • Aim 1 a: Determine the effects of prespecified clusters of MCCs on hospital readmission. • Aim 1 b: Develop machine learning algorithms to identify computergenerated clusters of MCCs associated with hospital readmission.
Challenges Training Infrastructure (Accessing resources and developing a team)
Training Opportunities & Resources Short Courses and Workshops • + Data Science includes online and in-person learning • Machine Learning Summer & Winter School • Machine Learning for the School of Medicine Formal Programs • Master in Interdisciplinary Data Science › Clinical mentor and/or capstone partner • Duke University Professional Studies Online Program in Big Data and Data Science Other Resources • Duke Forge • Machine Learning community at Duke › ML Day
Training Opportunities & Resources NIH-funded Centers, Workshops, and Fellowships • • • Big Data 2 Knowledge Data Science @ NIH NINR’s Methodologies Boot Camp Other Resources • • • NC State Data Matters SAS Academy for Data Science (Big Data + Analytics) UNC Coding and Data Analytics Bootcamp
Build an Interdisciplinary Team Principal Investigator CTSA Biostats Core Sponsor ML Expert $$ IT/Database Admin Data Manager/ Programmer Statistician
Lessons Learned Along the Way 1. Secure sufficient funding 2. Understand basic differences between ML & Statisticians 3. Determine desired level of involvement and understanding 4. Build an Interdisciplinary Team 5. Assess IT infrastructure and systems
Acknowledgements Duke Clinical & Translational Science Institute – CTSA PI, L. Ebony Boulware, MD, MPH – CTSA KL 2 Leaders: Laura Svetkey, MD, Kimberly Johnson, MD, MHS – Rasheed Gbadegesin, MD, MBBS CTSA KL 2 Mentoring Team/Collaborators – – – – – Cathleen Colón-Emeric, MD, MHS Elizabeth Merwin, Ph. D, RN Wei Pan, Ph. D Qing Yang, Ph. D Sathya Amarasekara, MS Lesley Curtis, Ph. D Lawrence Carrin, Ph. D Sayan Mukherjee, Ph. D Rachel Drelos, MD/Ph. D student Duke University School of Nursing
Michael P. Cary, Jr. , Ph. D, RN Associate Professor Duke University School of Nursing Office: Pearson Building, Room 3134 307 Trent Drive, Durham, NC 27710 Phone: (919) 613 -6031 Email: michael. [email protected] edu