Misspelled Health Queries Are Common, But AI Can Fix Them

When people search for health information online, they often type quickly on mobile devices, leading to misspelled queries that can block access to vital medical answers. A new study provides the first rigorous measurement of this problem, revealing that 61.5% of real consumer health questions contain at least one spelling error, with an average of 11% of words in these queries misspelled. This high error rate, derived from datasets like the TREC 2017 LiveQA Medical track with 104 user-submitted questions, highlights a pervasive barrier in healthcare question-answering systems where a single typo can prevent patients from finding relevant information about conditions or medications. The research underscores the urgent need for practical solutions in real-world applications, as these errors are far more common than in professional medical texts.

The key finding from the study is that correcting these misspelled queries significantly improves the quality of search . Using s like edit distance correction, the researchers achieved a 9.2% improvement in Mean Reciprocal Rank (MRR) and an 8.3% boost in NDCG@10 scores over uncorrected baselines when using BM25 retrieval on 1,935 MedQuAD answer passages. This means that after correction, relevant answers were ranked higher and found more consistently, with MRR increasing from 0.633 to 0.691. Importantly, the study found that correcting only the answer corpus—without fixing the queries—provided minimal benefit, with just a 0.5% MRR improvement, confirming that the error lies primarily on the user side.

To conduct this analysis, the researchers employed a controlled ology using real, naturally occurring spelling errors from two public datasets: TREC LiveQA Medical and HealthSearchQA. They built a domain vocabulary of 8,201 terms from the MedQuAD corpus and evaluated four correction s: conservative edit distance, standard edit distance, context-aware ranking, and SymSpell. In experiments, they tested conditions such as uncorrected queries against an uncorrected corpus, uncorrected queries against a corrected corpus, and fully corrected queries against a corrected corpus, using retrieval metrics like Recall and MRR based on human relevance judgments.

Showed that query correction drove almost all the observed improvements, with edit distance s modifying 68.9% of queries and achieving the highest gains. For example, in BM25 retrieval, conservative edit distance improved MRR by 7.9%, while SymSpell underperformed with only a 1.3% MRR increase due to its delete-based approach being less suited to medical vocabulary. An error analysis of 100 samples per revealed that 70-86% of corrections were partial improvements, meaning they changed tokens but not always to the exact intended word, yet still enhanced retrieval. The data also indicated that corpus correction alone could slightly degrade TF-IDF retrieval, cautioning against this approach in production systems.

These have clear for healthcare systems, suggesting that implementing query-level spelling correction is a low-cost, high-value intervention that can be deployed with minimal latency. The study recommends conservative edit distance as the default for safety-critical applications, as it balances accuracy with reduced risk of harmful corrections, such as confusing 'hypertension' with 'hypotension'. However, limitations include the small query set of 103 evaluated questions and the use of TF-IDF as a proxy for dense retrieval, which may not fully capture the behavior of modern neural models like PubMedBERT. Future work should explore LLM-based correction and evaluate confusable term safety layers to further enhance real-world healthcare QA systems.