Electronic patient records contain a wide range of diverse information, which has significance in regards to patient follow-up and further treatment. These records consist of vast amounts of data, much of which is in free text form, which can be time consuming for the clinicians to read tho through, in search of important diagnoses. It is therefore necessary to develop documentation systems that support these large amounts of data. If this data can be structured, the benefits include time conservation for the clinician, and quality of treatment for the patient.
The goal of this project was to investigate how neural networks can be used to detect diagnoses and their context in ECG notes, when taking into account that the amount of annotated data for clinical NLP is limited.
Two LSTM networks were used for the detection of candidate concepts and context classification, respectively.
In context classification, on a subset of ECG notes, an F1 score of 0.92 was measured.
The system was also tested on the shARe dataset, where the concepts were linked to a SNOMED CT code. A precision, recall and strict F-score of 0.98, 0.98 and 0.98 was obtained, respectively. The model was found to be capable of detecting diagnoses with a high level of confidence.