Natural Language Processing and Machine Learning Methods to Characterize Unstructured Patient-Reported Outcomes:...

Area of Science:

• Computational linguistics
• Medical informatics
• Oncology

Background:

• Patient-reported outcomes (PROs) are crucial for assessing survivorship in pediatric cancer patients.
• Traditional PRO surveys can be supplemented by analyzing unstructured data from clinical interviews.

Purpose of the Study:

• To validate natural language processing (NLP) and machine learning (ML) algorithms for identifying pain interference and fatigue attributes in child and adolescent cancer survivors.
• To compare the performance of NLP/ML algorithms against expert-adjudicated PRO data.

Main Methods:

• A cross-sectional study involving 391 meaning units for pain interference and 423 for fatigue from interviews with pediatric cancer survivors (aged 8-17) and caregivers.
• Semantic labeling of meaning units by content experts into physical, cognitive, and social attributes.
• Validation of two NLP/ML methods: bidirectional encoder representations from transformers (BERT) and Word2vec with support vector machine or extreme gradient boosting.

Main Results:

• BERT demonstrated superior accuracy in identifying cognitive and social attributes of pain interference and fatigue compared to Word2vec-based methods.
• BERT achieved high accuracy scores (e.g., 0.931 for pain interference cognitive attributes) and superior areas under the receiver operating characteristic and precision-recall curves.

Conclusions:

• The BERT model shows high validity and accuracy for assessing patient-reported outcomes in pediatric cancer survivors.
• NLP/ML methods, especially BERT, provide a viable and effective alternative to standard PRO surveys for capturing symptom experiences in clinical settings.