OBJECTIVES:

Literature analysis could benefit from machine-learning (ML) methods that parse medical text to extract reported results. We demonstrate this by creating an algorithm for automated discovery of Comparative Effectiveness Hypotheses for medical interventions.

METHODS:

We ran an ML algorithm, using Text Classifiers and Named Entity Recognizers, against a corpus of PubMed abstracts. The algorithm identified the study type (e.g., Randomized Control Trial), and parsed diseases and conditions, interventions, group-sizes and fractional outcomes from the abstract text (e.g., the phrase, "RA clinical remission was 6 of 8 for infliximab" yields {condition: rheumatoid arthritis; intervention: infliximab; outcome: clinical remission; outcome-count: 6; group-size: 8}). Text-spans identifying conditions, interventions, and outcomes were normalized across documents using clustering and ontology alignment.

The extracted data was used to construct a table containing Relative Risk (RR) values with 95% CI. For each intervention, paired conditions and outcomes are found, and the RR values are calculated from the fractional outcome results of the intervention and its competitors, aggregated across studies. Over-/Under-Perform Hypotheses are those condition/outcome pairs for which RR=1.0 falls outside the 95% CI. Statistics were calculated over a random sample of interventions with >= 100 published abstracts.

RESULTS:

The average number of generated hypotheses for each intervention was 4.0, 3.6, and 10.2 for over-perform, under-perform, and no-significant-difference. Average counts of disease conditions for each intervention were 3.5, 3.2 and 7.5, respectively.

CONCLUSIONS:

We demonstrated that AI can be used to assist in discovery of comparative effectiveness hypotheses from medical literature. Generated hypotheses spanned a range of conditions and outcomes for each intervention.