THE MALONE CENTER FOR ENGINEERING IN HEALTHCARE The Role of Data in Achieving Precision and Value in Healthcare Gregory D. Hager Mandell Bellmore Professor Director
The Malone Center Mission: Transform the Process of Healthcare Delivery By Translating Research-Based Innovations into Engineered Systems NAE / IOM (2006) IOM (2011) PCAST (2014) Advancements in medicine and technology are only as effective as the associated care delivery M A L O N E C E N T E R . J H U . E D U
Health Care: Some Numbers o Total Expenditures: 3T (18% of GDP) Value = Quality/Cost o System inefficiencies estimated to be up to 1/3 cost o Medical errors estimated to be 3 rd leading cause of death in the US Makary & Daniel. "Medical error-the third leading cause of death in the US." BMJ: British Medical Journal, 2016. Patient workflow complexity Decision complexity from Basole et al. J Am Med Inform Assoc 2015.. from Engineering a Learning Healthcare System, NAM M A L O N E C E N T E R . J H U . E D U
Opportunity: A New Wealth of Data 2300 exabytes of healthcare data will be produced in 2020 (153 in 2013) o 5,320,357 patients in Epic o 35,986,859 notes written in Epic in 2017 o 12,914 average patient encounters documented in Epic per day o 5,000+ physicians in Epic o 2,783,734,072 DICOM objects, 12,140,267 studies for 2,421,774 patients From 2008 to 2014, hospitals with EHRs rose to 75% from 9%, and in doctors’ offices rose to 51% from 17%. M A L O N E C E N T E R . J H U . E D U
Data Science Opportunities at Multiple Levels 4-Level Health Care System Monitoring/Modeling Clinical decision making Hospital operations Public policy Data Science Opportunities toward better: health outcomes • care value = quality / cost • Ferlie and Shortell, 2001 M A L O N E C E N T E R . J H U . E D U Slide courtesy Scott Levin
Opportunity Spaces o Diagnosis o Better decision-making o Early warning systems o Disentangling multiple causal factors E-Triage, Levin, JHU o Treatment o Choosing the most effective care o Monitoring quality of care o Improving training and workflow Schleroderma, Saria, JHU o Recovery o Better monitoring o Better analytics o Technology support beyond CCE Kata Project, Ahmad, JHU M A L O N E C E N T E R . J H U . E D U
A Challenge: Failed Surgery 0.25M sinus surgeries per year $22B per year expenditure on chronic rhinosinusitis 25% of surgeries for nasal airway obstruction “failed” What causes nasal obstruction? When will surgery help? Hypothesis: Large population data sets, correlated with outcomes can provide clues to who benefits and why M A L O N E C E N T E R . J H U . E D U
A Challenge: Failed Surgery 0.5M spinal fusion surgeries / yr $12B / year (70% increase 2001 – 2011) 7.5% compound annual growth by 2019 High Range in Variability (Quality) 53% of patients have comorbidity 8-25% of patients rehospitalized High variability in surgical outcomes à Opportunity for a more information-driven approach Courtesy Jeff Siewerdsen (JHU BME) M A L O N E C E N T E R . J H U . E D U
An Opportunity for “Image Clouds” Courtesy Jeff Siewerdsen (BME) M A L O N E C E N T E R . J H U . E D U
Statistical shape models Deformable registration à population shape change à statistical model ⋯ v ## v &# v ( * # v ( * & v #& v && ⋯ " # = " & = " ( * = ⋮ ⋮ ⋮ v #( ) v &( ) v ( * ( ) A Sinha , et al., Automatic segmentation and statistical shape modeling of the paranasal sinuses to estimate natural variations , SPIE Medical Imaging, 2016 A Sinha , et al., Simultaneous segmentation and correspondence improvement using statistical modes , SPIE Medical Imaging, 2017 12 M A L O N E C E N T E R . J H U . E D U
Variance along the principal mode for the maxillary sinus Statistical shape models Front view Left view Variance along the principal mode for the middle turbinates M A L O N E C E N T E R . J H U . E D U
Studying Effect of Anatomy on Outcomes What causes nasal obstruction? When will surgery help? Shape Space ⋯ s 1 s 2 s n M A L O N E C E N T E R . J H U . E D U
Studying Effect of Anatomy on Outcomes What causes nasal obstruction? When will surgery help? s 1 s 2 Shape Space ⋯ s 1 s 2 s n M A L O N E C E N T E R . J H U . E D U
Studying Effect of Anatomy on Outcomes What causes nasal obstruction? When will surgery help? s 1 s 2 Shape Space ⋯ s 1 s 2 s n M A L O N E C E N T E R . J H U . E D U
Studying Effect of Anatomy on Outcomes What causes nasal obstruction? When will surgery help? s 1 s 2 Shape Space ⋯ s 1 s 2 s n M A L O N E C E N T E R . J H U . E D U
Data-Driven Anatomic Models From Endoscopy Structure from Motion + Deep Learning 3D surfaces and normals Deformable Registration Statistical Model Move from qualitative assessment to quantitative measurement of anatomy of every patient With R.H. Taylor, A. Sinha, A Reiter, M Ishii Patient Model M A L O N E C E N T E R . J H U . E D U
The Human Element: Skill vs. Outcomes Score Bottom Quartile 6.30% Readmission 2.70% Score Top Quartile 3.40% Reoperation 1.60% 14.50% Complication 5.20% Michigan Bariatric Surgery Collaborative Samples : 0.26% 20 bariatric “expert” surgeons ranked by at Mortality 5x Mortality Rate! 0.05% least 10 reviewers. 10,343 patients admitted 2006-2012 Expertise Score 1 2 3 4 5 Birkmeyer J.D, et al. Surgical Skill and Complication Rates after Bariatric Surgery. NEJM, 2013. M A L O N E C E N T E R . J H U . E D U
Temporal Convolution Networks (TCN) Max over filter activations: Index of max filter: Layer 1 Layer 2 Layer 3 DiPietro, Robert, et al. "Recognizing surgical activities with recurrent neural networks." MICCAI , 2016. Lea, Colin, et al. "Temporal Convolutional Networks for Action Segmentation and Detection." CVPR . 2017. M A L O N E C E N T E R . J H U . E D U
A New Lens on The Human Element Novice Septoplasty Capturing and structuring surgical performance data and relating it to quality of outcome. Expert M A L O N E C E N T E R . J H U . E D U
First Steps Towards an AI-Assisted OR Classifying skill in the laboratory Classifying skill in the operating room Less experienced operator Resident More experienced operator Attending 1 Linear classifier Linear classifier − − 2.4 0 − − 2.6 − 1 − − 2 − − − 2.8 Similarity to Expert Similarity to Expert − 3 − − − 3 − 4 − − 5 − − − 3.2 − 6 − − − 3.4 − 7 − − 8 − − − 3.6 − 9 − − 8 − 7 − 6 − 5 − 4 − 3 − 2 − − − − − − − − − − − − − − − − − 5 − 4.9 − 4.8 − 4.7 − 4.6 − 4.5 − 4.4 − 4.3 − 4.2 Similarity to Novice Similarity to Novice Trained in the lab Tested in the OR M A L O N E C E N T E R . J H U . E D U
Following Treatment and Recovery A Plug and Play platform for quantifying clinical activities Identifying Objects Identifying People Table Caregiver Patient Bed [upright ] Standing Sitting Detecting People Spatio-Temporal Motion Analysis Estimating Pose Analysis Ma, Rawat, Reiter et al. Crit. Care Med., 45:4, 2017 M A L O N E C E N T E R . J H U . E D U
Following Treatment and Recovery o Demonstration: Assessing patient mobility Estimating Pose Detecting People Standing Sitting Patient Caregiver Motion Analysis Spatio-Temporal Analysis Identifying People Ma, Rawat, Reiter et al. Crit. Care Med., 45:4, 2017 M A L O N E C E N T E R . J H U . E D U
Early Results Physician Nothing In-bed Out-of-bed Walking Nothing 18 (22%) 4 (5%) 0 0 In-bed 3 (4%) 25 (30%) 2 (2%) 0 System Out-of-bed 0 1 (1%) 25 (30%) 1 (1%) Walking 0 0 0 4 (5%) Total 21 (26%) 30 (36%) 27 (32%) 5 (6%) § Weighted Kappa: 0.86 (95% Confidence Interval: 0.72, 1.00) § Sensor and clinician agreed on 72 out of 83 segments (87%) § Of the 11 discrepancies, 7 were due to confusion between ‘nothing in bed’ and ’in-bed activity’. 1200+ hours of data collected in the Johns Hopkins Weinberg ICU Ma, Rawat, Reiter et al. Crit. Care Med., 45:4, 2017 M A L O N E C E N T E R . J H U . E D U
Summary What you cannot measure, you cannot improve – Lord Kelvin o Healthcare is being transformed by data o We are still in the early days of learning how to effectively collect and use data to improve value o Data science challenges are large and growing: bias, missing data, heterogeneity and robustness, transparency, trust, and privacy o Challenges remain at the first (data acquisition) and last (deployment) mile. M A L O N E C E N T E R . J H U . E D U
Recommend
More recommend
