mentored (MEtadata aNnotaTiOn pREDiction) is a BiLSTM neural network trained to predict UMLS Codes to a given metadata description. The network consists of three layers: embedding layer, BiLSTM with 2 cells and a linear classifier.
| Model | t1 | t3 | t5 | t10 |
|---|---|---|---|---|
| 2BiLSTM w/ Augmentation | 67.27 | 72.85 | 74.23 | 75.63 |
| Phrase | 1. Prediction | 2. Prediction | 3. Prediction |
|---|---|---|---|
| Name des Patienten (name of the patient) | Patient surname (25.04%) | Medication name (18.32%) | Patient forename (17.45%) |
| Krebs der Niere (cancer of the kidndy) | Kidney cancer (24.84%) | Subject Diary (17.16%) | Malignant neoplasm of kidney (16.12%) |
| Krebs der Leber (cancer of the liver) | Malignant Placental Neoplasm (21.20%) | Secondary malignant neoplasm of liver (18.50%) | Liver reconstruction (17.72 %) |
| Blut (blood) | Blood (19.38%) | Blood in Urine (17.81%) | Coagulation Process (17.21%) |
You need the data files in order to run the training. Two files are needed: MDM metadata and the corresponding MeSH files. Both files need to have the following shapes: "CODE";"PHRASE" e.g. "C0027989;newspapers"
Put the files in the data directory, change to path in the file "start_training" and run!