Predict_EMR

Question 1

Potential uses - “Developing Predictive Models Using Electronic Medical Records: Challenges and Pitfalls”, Paxton et al

Answer 1

• risk stratification or early detection (*)
• biomarker discovery
• cohort detection for clinical trials
• optimization of resources usage

Question 2

EMR data collection: pros and cons

Answer 2

Pros
- virtually zero-cost and abundant
- near real-time
Cons
- incidence of the condition tracked (how often it occurs)
- privacy

Question 3

Predict with EMR: key issues discussed

Answer 3

• confounding medical interventions (CMI) can mask the ground truth labels used for training/evaluation
• importance of clinical use and its effect on training and evaluating the predictive models

Question 4

Clinical applications evaluated

Answer 4

• standalone monitoring/diagnostic - less likely w/ EMR data

- assisted monitoring/alert (*) - more suitable w/ EMR data

Question 5

Case study: septic shock using the MIMIC-II EMR database

Answer 5

• each patient = set of time points, w/ time point an entry in EMR system
• more measurements are taken by personnel during periods of interest
• checked if values fall within an acceptable range in order to account for errors in data entry
• pressors or antibiotics are used as indicators of treatment
• for early prediction system will discard the time points after treatment has begun
• used a subset of 77 chart variables: a combination of chart data, lab results, patient demographics, fluid I/O events, and medications
• tracked variablilty using maximum, variance, least-squares fit line over the last 6, 12, and 24 hours
• added more tracking data to a total of 1011 features

Question 6

Challenges

Answer 6

• incomplete observations: transferred too late, discharged too early (discard data w/ less than 12h prior to sepsis)
• selection bias: nature of the practive, care unit, geographical location, demographics, health condition at admission
• confounding medical interventions: medical interventions that will affect the risk of the outcome of interest: i.e. cannot make the diff between right treatment or conservative treatment
• the paper defines ‘clear’ and ‘confounded’ data samples w/ ‘confounded’ = CMI applied and no adverse reaction is observed

Question 7

Assisted monitoring definition

Answer 7

• the system is trained to alert only on ‘at risk’ patients w/ or w/o CMI where adverse condition is observed, thus using only ‘clear data’ samples;

Question 8

Model data settings

Answer 8

• data is considered in sliding windows of 72h; for septic shock only one window of 72h is considered with data prior to that being discarded
• a particular patient’s data is included in either the training set or test set
• classes are balanced by up-weighting the minority class by the ratio positive to negative examples

Question 9

Models by CMI handling

Answer 9

uses SVM-light with a linear kernel and default parameters to train a predictive model for each of the four approaches, and evaluate their performance in the context of an assisted monitoring system:

• with clear data only
• with CMI as positive samples
• with CMI as negative samples
• with CMI unknown

Question 10

Error sources

Answer 10

To understand the utility of each model (…) performed errror analysis done by the severity of the adverse condition

• many mistakes on patients that develop severe sepsis but do not progress to septic shock
• evaluate the discriminative power only on patients with septic shock against patients that do not progress beyond SIRS (mild sepsis) given that misclassifications of severe sepsis patients should not be considered as an error