Data quality requirements
For ComplianceCoach to provide predictions to prioritize HME outreach, the data must satisfy the following:
Completeness: To limit the rate of false positives, the model training and prediction data must not be missing more than 30% of the data points. We set the 30% missing data point rate threshold after simulating the effect of randomly missing data points in the dataset.
Note: Patients on card-to-cloud devices may be missing data for extended periods. For these patients, ComplianceCoach displays an SD card icon in the patient list and inserts an N/A.
Accuracy: Device usage in AirView can exceed 24 hours (occurrence is around one in every million data points). For these instances, the model clips the value so that device usage fed into the model is between 0 and 24 hours.
Timeliness: The delay between AirView data in our databases for patients should not consistently exceed more than 24 hours to ensure that predictions displayed in ComplianceCoach are up to date. HMEs can quickly determine the most recent prediction date via ComplianceCoach. If data fails to update after 48 hours, HMEs will see an info icon warning in the patient list and the last available prediction and prediction data in the Forecast tab.
Validity: The device usage features must be floating-point values, and the number of compliant days and days since setup must be whole numbers for all values in the dataset.