Machine Learning for Healthcare

Reading Questions

Session 3 Reading Questions

• What are some of the challenges of using retrospective data from electronic health records?
• The paper states that “the presence of a white blood cell count test order is associated with a 2.1% lower survival rate (P<0.001).” What does the P<0.001 signify?
• All else equal, would we expect the survival rate to be higher for a patient with a 5pm white blood test or a 5am white blood test?
• All else equal, would we expect the survival rate to be higher for a patient with a Sunday white blood test or a Wednesday white blood test?
• Besides time of lab test and the lab test value, what other healthcare process variables might give additional signal for prediction of 3-year survival?

Session 4 Reading Questions

• [Razavian et al, 2015] The trained model used in this paper might not identify high risk patients well for a rural hospital in Wyoming. Name two potential reasons why.
• [Razavian et al, 2015] In Figure 1, we see that Patient G is excluded because they experienced diabetes onset during the gap period. Why is there a gap period for this prediction task?
• [Pozen et al, 1984] Based on the paper’s results, name three ways we would expect the CCU to change if we rolled out the model completely and physicians followed the model’s recommendations.

Session 5 Reading Questions

• [Caruana et al.] What are the different approaches that the authors suggest for interpreting the results from learning with few versus very many features, and why are the different approaches needed?
• [Caruana et al.] What utility does introducing a bias term serve in terms of improving the interpretability of the learned model?
• [Futoma et al.] Before their “real-time validation” experiment, how do they decide how much data to consider for a negative example (i.e. one who never develops sepsis)?

Session 6 Reading Questions

• [Quinn et al, 2009] What problem is the factorial model trying to solve that the SLDS model struggles with?
• [Hannun et al, 2016] How does the performance of the DNN compare to the aggregated cardiologist score? How do the mistakes made by the DNN compare to those made by the cardiologists?

Session 7 Reading Questions

• [Leaman et al, 2015] What makes identifying disorder mentions in clinical text difficult?

Session 10 Reading Questions

• What is Echocardiography?
• What is one reason why a doctor might want a VNAP to provide versioning?

Session 11 Reading Questions

• What are the three models that the authors compare to build a health knowledge graph? (Give names of models, no need to elaborate.)
• How are diseases and symptoms extracted from the clinical records?
• Which symptom did the authors omit from the analysis?

Session 12 Reading Questions

• What are patches and how are they used in the model outlined by the authors?
• How did the authors incorporate a clinician into the pipeline?

Session 13 Reading Questions

• What are sensitivity and specificity? Overall how do CAD systems and double-readings compare on sensitivity and specificity?
• What are two challenges from using patch-centered methodologies?
• What is tomosynthesis and how does it compare to traditional mammography?

Session 14 Reading Questions

• Examine Table 1.1. Assume our universe only includes gods whose names begin with “H”. This leaves us with 6 gods. What is P(Y^(a=0) = 1 | “H” gods)? What is P(Y^(a=1) = 1 | “H” gods)?
• Examine Table 1.2. Assume our universe only includes gods whose names begin with “H”. This leaves us with 6 gods. What is P(Y=1 | A=0, “H” gods)? What is P(Y=1 | A=1, “H” gods)?
• Explain “confounding” in 15 words or less. Does not have to be a complete sentence.

Session 16 Reading Questions

• The authors seek to model mechanical ventilation in the ICU as a off-policy reinforcement learning algorithm. What is the action space? What is the reward space? (Short answers okay!)
• What is the main difference between on-policy and off-policy reinforcement learning? (can be short!) Why do the authors choose to use off-policy learning here?
• Figure 6 shows the learned feature importance. Pick one value and explain whether or not it makes sense to be a high/low weight.

Session 17 Reading Questions

• One difference between Komorowski et al 2018 and Prasad et al 2017 is the difference in reward function. What is the reward function here?
• Name two limitations of the study.

Session 22 Reading Questions

• Draft a paragraph on component or concepts in the white paper that the agency should:
  1. Keep in future iterations
  2. Remove or clarify in future iterations
  3. Add (for missing components)Suggestions: consider topics covered in previous lectures including generalizability, evaluation, updating models, fairness, causality, etc.