A Disease Identification Algorithm for Medical Crowdfunding Campaigns: Validation Study.

Area of Science:

• Computational linguistics
• Health informatics
• Medical sociology

Background:

• Web-based crowdfunding is increasingly used for medical expenses, generating valuable but unstructured data.
• Analyzing this text data for specific medical conditions presents significant research challenges.
• Existing methods struggle with scalability and accuracy for large datasets.

Purpose of the Study:

• To validate an algorithm for identifying 11 disease categories in medical crowdfunding campaigns.
• To combine Named Entity Recognition (NER) and keyword searching for improved disease identification.
• To facilitate large-scale research on health crowdfunding data.

Main Methods:

• Web scraping collected 89,645 GoFundMe campaigns.
• A custom algorithm used pretrained Spark NLP for Healthcare models (NER, entity resolution) and keyword searches.
• Conditions were mapped to International Classification of Diseases, 10th Revision, Clinical Modification (ICD-10-CM) codes.
• Algorithm performance was assessed against 400 manually labeled campaigns.

Main Results:

• The algorithm achieved high interrater reliability (Cohen κ: 0.69-0.96) for disease category detection.
• NER identified 6,594 unique ICD-10-CM codes; keyword search added 3,261 more campaigns.
• Overall algorithm performance: precision 0.83, recall 0.77, F1-score 0.78, accuracy 95%.
• Performance varied by disease category but remained high across the board.

Conclusions:

• A novel algorithm effectively identifies 11 disease categories in medical crowdfunding text.
• The approach combines NLP and ICD-10-CM coding for high precision and accuracy.
• This method enables robust research into health conditions within crowdfunding data.