T21. Predicting Pulmonary Air Leak Resolution Using Transpleural Airflow Data after Lung Resection
*Sebastien Gilbert1, Daniel G. French2, Natalie Japkowicz3, Mohsen Ghazel1
1University of Ottawa, Ottawa, ON, Canada; 2Dalhousie University, Halifax, NS, Canada; 3American University, Washington, DC
Objective: To develop an on-demand forecasting model in order to determine the optimal timing to remove chest tubes after pulmonary resection using digitally recorded airflow measurements.
Methods: Digitally recorded airflow data from 32 patients who underwent pulmonary resection was used to estimate a statistical time-series forecasting model of future airflow measurements using historical airflow data. An Autoregressive Integrated Moving Average (ARIMA) statistical forecasting model is fitted to the most recently recorded airflow measurements over a time frame of length Thistory hours. This model is then used to predict future values of the airflow over the immediate forecasting horizon of length Thorizon hours. The inherent non-stationarity in the airflow time series is avoided by limiting the historical data period to Thistory = 24 hours or less. Also, to account for the uncertainties associated time series predictions, we limit the forecasting horizon to Thorizon = 24 hours or less and generate many (1000) possible forecasting paths of potential predicted values. The likelihood of the safe removal of the chest tube is then estimated by the percentage of these forecasting paths, which satisfy the following criterion: predicted airflow does not exceed 30ml/min for the next 8 hours. In order to achieve more distant prediction and require the airflow signal to stabilize, we tested a more practical clinical scenario with a forecasting horizon of Thorizon = 24 hours and the chest-tube is assumed to be safely removed if the above criterion is met over the last 8 hours (i.e., between 16th and the 24th hour). The predicted times for removal of the chest tubes were then compared to the times when chest tubes were actually removed.
Results: The performance of the proposed system was evaluated on 32 patients (Table 1). With 95% probability, the system correctly predicted to maintain the chest tube for 7 patients and remove the chest tube for 20 patients, while it incorrectly predicted to maintain the chest for the remaining 5 patients (15.83% of 32 patients). For the 20 patient with correctly predicted chest tube removal, the system forecasted removal times that are significantly lower than the actual chest tube removal times, resulting in an average time saving of 3877 minutes and the 95% confidence interval of the saved time is (932.89, 6821.11) minutes. The system did not predict any chest tube removal at times when they should not have been removed, hence eliminating the need for re-insertion.
Conclusions: It is possible to predict the future airflow from a chest tube over the next 8, 16 and 24 hours using a previously measured values. Predicting airflow will facilitate earlier and timelier removal of chest tubes and enhance discharge planning, allocation of hospital resources and patient satisfaction. This prediction model can also prevent adverse events associated with premature removal of chest tubes.