Germination Project Data Science at Penn Medicine

1.

Data Science and Predictive Healthcare Corey Chivers, PhD Senior Data
Scientist @CjBayesian

2.

2 Outline wWhat is Data Science? wBecoming a data scientist
wTools of a data scientist wPredictive Healthcare at Penn Medicine

3.

3 wWhat is Data Science?

4.

4 wWhat is Data Science? wWhat do you need to
study to become a Data Scientist?

5.

5 Data Science Venn Diagram http://drewconway.com/zia/2013/3/26/the-data-science-venn-diagram

6.

6 Popular Expectations

8.

8 Using data and computation to give people superpowers

9.

9 One Data Scientist’s Story w Undergraduate Chemistry McGill University

10.

10 One Data Scientist’s Story w Undergraduate Chemistry Atmospheric Physics
McGill University

11.

Environmental Science McGill University

12.

Environmental Science w Non-Profit research & activism around public space use in Toronto McGill University

13.

Environmental Science w Non-Profit research & activism around public space use in Toronto w PhD Computational Biology McGill University

14.

Environmental Science w Non-Profit research & activism around public space use in Toronto w PhD Computational Biology Modeling population dynamics for conservation McGill University

15.

15

16.

16

17.

17

18.

18

19.

19 The medical record is full of data* *There’s a
whole bunch that is not in the EHR, too! There are all kinds of patterns in there! Provides predictions to aid in clinical decision making

20.

20 “Predictive Health Care” is the development and integration of
Data Science* into clinical and operational processes and workflows to deliver better outcomes at lower cost. * Machine Learning, data visualization, computational simulation, AI, etc.

21.

21 Cat Cat Cat Dog Dog Doge Machine Learning Generalizing
from many examples

22.

22 0.7 Dog 0.25 Cat 0.05 … Cat Dog Machine
Learning Generalizing from many examples

23.

23 Time Encounter with First Diagnosis of Valve Disease (‘ground
truth’ ICD code) Patient Encounters Diagnoses (non-valve disease) Labs Etc. Input data Time Input data Positive Case Negative Case No encounters with diagnosis of Valve Disease (‘ground truth’ ICD code) Clinical Example

24.

24 Penn Medicine Data Data Type 2011-Present estimated Total Vital
(Manual) 20GB Vital (Telemetry) 1.5TB/year Labs 36 GB Meds and Orders 42 GB Radiology Raw 250TB Radiology Meta 200 GB Clinical Notes 250 GB/year Twitter 100TB Other social media (est.) 1TB Wearables (est.) 5GB/year Genomics 800TB Lots of examples to generalize!!

25.

25 Data Visualization

26.


27.


28.

28 Computational Simulation Using randomization to predict trajectories of complex
dynamic systems Allows you to experiment with ‘what if’ scenarios and to select from alternative actions. ‘Writing down’ the hypothesized system dynamics also sharpens understanding and assumptions. (Chivers, 2014)

29.

29 w Creating ‘Digital Twins’ that allow a team to
explore efficiency gains • Build optimized scheduling templates • Test process changes to drive clinical workflow redesign • Identify bottlenecks w We’ve created digital twins to support Emergency Department and OB/GYN clinic operations Patient Volume Minute Census Wait Times OB/GYN Simulated Output Computational Simulation

30.

30 Tools

31.

31 Skills Successful Data Science is highly collaborative

32.

32 As a ____________ ROLE when I’m ____________ CONTEXT if
I knew ____________ INFORMATION I would do _____________ INTERVENTION to improve __________ OUTCOME Data Science “MadLibs” Solving the right problem

33.

33 Mechanical Ventilation

34.

34 If I knew when mechanical ventilator patients were ready
for a Spontaneous Breathing Trial (SBT) Mechanical Ventilation

35.

35 I would reduce sedation and initiate an SBT trial
earlier to decrease how much time patients spend in the ICU Mechanical Ventilation

36.

36 Time on Ventilation Mechanical Ventilation Time in ICU

37.

37 Lung Connect

38.

38 If I knew which lung cancer patients will go
to the ED Lung Connect

39.

39 I would address their symptoms in the outpatient setting
to lower ED usage Lung Connect

40.

40 flic.kr/p/aDbss E Lower ED Usage Lung Connect

41.

41 https://www.pennmedicine.org/news/news-blog/2018/june/palliative-connect-digitizing-the-physicians-intuition-to-prompt-critical-conversations Palliative Connect

42.

42 https://www.pennmedicine.org/news/news-blog/2018/june/palliative-connect-digitizing-the-physicians-intuition-to-prompt-critical-conversations If I knew which patients have serious, life-limiting
illnesses Palliative Connect

43.

43 I would make sure they receive a palliative care
consultation to ensure the care team understands their goals and desires https://www.pennmedicine.org/news/news-blog/2018/june/palliative-connect-digitizing-the-physicians-intuition-to-prompt-critical-conversations Palliative Connect

44.

44 Earlier access to palliative care More documented advanced care
plans Palliative Connect

45.

45 Some Challenges w Data != Reality • Healthcare data
is particularly challenging – Non-random missingness – Largely unstructured – Clinical concepts and their representations change over time w How good does the model need to be? • We use meta decision theory to decide whether the model makes better decisions than some alternative w How do we know it’s working? • When you’re trying to prevent the thing you’re predicting, is your prediction bad or your intervention good?

46.

46 Changing Clinical Concepts Weekly Count

47.

47 How good does the model need to be? Data
Scientist: Finally, I built a model from the data! Doc: Awesome, is it good? Data Scientist: I used tensorflow, so, ya. Doc: Lets deploy this thing!* *Any resemblance between these fictional characters and any persons, living or dead, is a purely coincidental

48.

48 How good does the model need to be? Data
Scientist: Finally, I built a model from the data! Doc: Awesome, is it good? Data Scientist: I used tensorflow, so, ya. Doc: Lets deploy this thing!* *Any resemblance between these fictional characters and any persons, living or dead, is a purely coincidental

49.

49 All models are wrong, but some are useful. -
George Box

50.

50 FP FN TP TN

51.

51 How good tho? How good tho? How bad tho?
How bad tho? FP FN TP TN

52.

52 Goodness can be measured in any units More adorbs
Less adorbs

53.

53 When outcomes are uncertain, the best decision is the
one that has the highest expected goodness.

54.

one that has the highest expected goodness. Machine Learning can only help you with this part!

55.

one that has the highest expected goodness. Machine Learning can only help you with this part!

56.

56 Healthcare Example Predicted Sepsis Treated a true case (Potential
to avoid bad outcome) Predicted Sepsis Treated a false case (unnecessary) Predicted No Sepsis Didn’t treat (all good) Predicted No Sepsis Failed to treat (Bad outcome)

57.

57 Healthcare Example

58.

58 Treat none Cost of Intervention ($) Cost of event
($) Treat all Pregnancy Related Hypertension (PRH) is the leading cause of maternal morbidity and mortality in the U.S. High-risk patients à remote blood pressure monitoring https://healthcareinnovation.upenn.edu/projects/heart-safe-motherhood

59.

59 Takeaways w Predictive Healthcare is the integration of Data
Science into clinical and operational workflows to improve healthcare quality w Data Science is a collaborative, multidisciplinary endeavor • Using data to make better decisions • There are many paths to becoming a data scientist! w Taking time to ensure you’re solving the right problem is essential

60.

60 Thanks!!

Germination Project Data Science at Penn Medicine

Germination Project Data Science at Penn Medicine

Want more?

Transcript