Click here to


Are you sure ?

Yes, do it No, cancel

An NLP-Based Collaborative Intelligence Tool to Support Precision Medicine

S Dieterich1*, P Arora2, S Azghadi1, (1) UC Davis Medical Center, Sacramento, CA, (2) UC Davis, Davis, CA


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Lack of access to peer-reviewed clinical trial publications and other resource constraints currently lead to multi-year delays in clinically implementing recommendation from newly published clinical trials, leading to cancer patients not receiving necessary radiation. Our work aims to build a collaborative intelligence physician support tool based on natural language processing techniques.

Methods: Because of prolific clinical trial data and mostly unambiguous treatment recommendations, early stage breast cancer was chosen as test scenario. A user interface, predictive literature search model, and automated data gathering subsystem were built. User-rating of suggested papers was implemented to improve modeling. Papers were evaluated based on similarity through term-frequency and inverse-document frequency analytics. Heuristic similarities were used to augment TD-IDF scores. We used Natural Language Processing techniques on abstracts from a subset of papers for primary component analysis with the goal to build similarity clusters to improve the model.

Results: The initial deployment of the model-based literature search component showed that the ML algorithm needs to be augmented by NLP-based clustering analysis to improve filtering of relevant information. The accuracy of the NLP analysis is limited by paywall restrictions of peer-reviewed publications, which currently limits utility of the decision support tool for physicians with limited subscription resources. Preprint publication of clinical trial data could overcome this barrier.
Principal component analysis suggests that papers cluster around treatment techniques (IORT vs. IMRT/VMAT), which will allow model finetuning based on the technology available to the physician user of the decision model.

Conclusion: This feasibility study of a decision support system has demonstrated that machine-learning algorithms combined with natural language processing can provide clinical radiation oncologists with a relevant selection of peer-reviewed clinical trial publications tailored to an individual patient’s pathology and staging. This reduces mental load and could shorten current multi-year delays of implementing evidence-based recommendations clinical practice.

Funding Support, Disclosures, and Conflict of Interest: SD has a research agreement with Varian.


Clinical Trials


Not Applicable / None Entered.

Contact Email