Background New drug treatments are regularly approved, and it is challenging to remain up-to-date in this rapidly changing environment. Fast and accurate visualization is important to allow a global understanding of the drug market. Automation of this information extraction provides a helpful starting point for the subject matter expert, helps to mitigate human errors, and saves time. Objective We aimed to semiautomate disease population extraction from the free text of oncology drug approval descriptions from the BioMedTracker database for 6 selected drug targets. More specifically, we intended to extract (1) line of therapy, (2) stage of cancer of the patient population described in the approval, and (3) the clinical trials that provide evidence for the approval. We aimed to use these results in downstream applications, aiding the searchability of relevant content against related drug project sources. Methods We fine-tuned a state-of-the-art deep learning model, Bidirectional Encoder Representations from Transformers, for each of the 3 desired outputs. We independently applied rule-based text mining approaches. We compared the performances of deep learning and rule-based approaches and selected the best method, which was then applied to new entries. The results were manually curated by a subject matter expert and then used to train new models. Results The training data set is currently small (433 entries) and will enlarge over time when new approval descriptions become available or if a choice is made to take another drug target into account. The deep learning models achieved 61% and 56% 5-fold cross-validated accuracies for line of therapy and stage of cancer, respectively, which were treated as classification tasks. Trial identification is treated as a named entity recognition task, and the 5-fold cross-validated F1-score is currently 87%. Although the scores of the classification tasks could seem low, the models comprise 5 classes each, and such scores are a marked improvement when compared to random classification. Moreover, we expect improved performance as the input data set grows, since deep learning models need to be trained on a large enough amount of data to be able to learn the task they are taught. The rule-based approach achieved 60% and 74% 5-fold cross-validated accuracies for line of therapy and stage of cancer, respectively. No attempt was made to define a rule-based approach for trial identification. Conclusions We developed a natural language processing algorithm that is currently assisting subject matter experts in disease population extraction, which supports health authority approvals. This algorithm achieves semiautomation, enabling subject matter experts to leverage the results for deeper analysis and to accelerate information retrieval in a crowded clinical environment such as oncology.

BACKGROUND New drug treatments are regularly approved and it is challenging to remain up-to-date in this rapidly changing environment. A fast and accurate understanding is important to allow a global understanding of the drug market; automation of this information extraction provides a helpful starting point for the subject matter expert, helps to mitigate human errors, and saves time. OBJECTIVE We apply NLP methods to classify disease populations within the free text of oncology drug approval descriptions from the BioMedTracker database, and extract the clinical trial entities that provide evidence for these approvals. METHODS We fine-tune a BERT model. This methodology has demonstrated state of the art results on a wide variety of NLP tasks. Therefore, we also expect it to be stable or improve over time as we increase the amount of input data. BERT’s performance is validated against a rule-based text mining approach. RESULTS By utilizing our fine-tuned BERT models, we achieve 61% and 56% 5-fold cross-validated accuracies for the line of therapy and stage of cancer classification tasks, respectively; with five classes each, this is a marked increase when compared to random classification. For the trial identification named entity recognition (NER) task, the 5-fold cross-validated F1 score is currently 87%. The training dataset is small (~400 entries) and both classification and NER task scores are expected to improve over time with the availability of additional data. For clinical validation of the model, the results were corrected by a subject matter expert before usage. The subject matter expert leveraged the results for further analysis as a helpful starting point in a crowded clinical environment such as oncology. CONCLUSIONS We developed a NLP algorithm that is currently assisting subject matter experts to extract stage of cancer, line of therapy and the relevant clinical trials that support these Health Authority approvals, from a free, unstructured text source. The increased structure these results bring can be further utilized in downstream applications, aiding searchability of relevant content against related drug project sources.