P0013 - Comparison of Various Machine Learning Models to Identify Peripherally Inserted Central Catheter (PICC) Tip Position from Radiology Reports

Background: Infants in the neonatal intensive care unit (NICU) frequently need peripherally inserted central catheters (PICC) to provide medications, parenteral nutrition, and fluids. An upper extremity PICC line tip is optimally positioned “centrally” in the superior vena cava (SVC) or at the SVC-right atrial (RA) junction. Malpositioned PICC lines run the risk of extravasation into surrounding spaces or causing vascular damage due to highly osmolar fluids. PICC line tip position must be confirmed frequently, as the tip may move from its original position, especially for small neonates. Automated detection of PICC line tip position may help alert bedside clinicians to non-central PICCs both during and after placement and may provide a helpful starting point for radiologists. This research seeks to use supervised machine learning to train and test an algorithm based primarily on text analysis of X-Ray reports from infants with an upper extremity PICC. Our hypothesis is that we will be able to automatically extract the radiologist-specified location of the PICC line tip from neonatal chest X-Ray reports with at least 98% accuracy.

Methods: A total of 17,596 chest and chest/abdominal x-ray reports met the criteria for inclusion. Based on the radiologist’s textual report, positions of the PICC lines were manually labeled with one of twelve separate anatomical locations: Cephalic, Axillary, Subclavian, Brachiocephalic, Subclavian/Brachiocephalic, Brachiocephalic/SVC, SVC, SVC/RA, RA, Neck, Chest, and Other. Boolean Accuracy was defined as a centrally located PICC tip in the SVC or SVC/RA. Once categorized, the data were split per category to maintain similarities between the training and testing data so that for each section approximately 30% of reports were randomly split into the test set, and the remaining 70% remained in the training set, and we benchmarked several machine learning algorithms.

Results: Resulting algorithms which yielded accuracies as high as 98% in identifying PICC lines in central vs. non-central position (Boolean Accuracy) and accuracies as high as 96% when attempting to categorize the individual anatomical location (12-category Accuracy). Please see the attached table.

Conclusions: Our study shows that machine learning models can automatically extract the anatomical location of PICC tips from radiology reports. We plan to implement this machine learning algorithm in the neonatal intensive care unit as a clinical decision support system in order to keep track of all central and non-central PICCs in the units. This approach shows promise in that it could be used to extract other types of data from radiology reports. Our next research task is to create an image recognition algorithm that will extract the PICC tip location based on the raw radiograph, and we will use our current algorithm to label additional training images for this supervised learning task.