{"id":34672,"date":"2025-07-12T12:27:16","date_gmt":"2025-07-12T10:27:16","guid":{"rendered":"https:\/\/www.graviton.at\/letterswaplibrary\/we-built-an-open-source-medical-triage-benchmark\/"},"modified":"2025-07-12T12:27:16","modified_gmt":"2025-07-12T10:27:16","slug":"we-built-an-open-source-medical-triage-benchmark","status":"publish","type":"post","link":"https:\/\/www.graviton.at\/letterswaplibrary\/we-built-an-open-source-medical-triage-benchmark\/","title":{"rendered":"We Built An Open-source Medical Triage Benchmark"},"content":{"rendered":"<p><!-- SC_OFF --><\/p>\n<div class=\"md\">\n<p>Medical triage means determining whether symptoms require emergency care, urgent care, or can be managed with self-care. This matters because LLMs are increasingly becoming the &#8220;digital front door&#8221; for health concerns\u2014replacing the instinct to just Google it.<\/p>\n<p>Getting triage wrong can be dangerous (missed emergencies) or costly (unnecessary ER visits).<\/p>\n<p>We&#8217;ve open-sourced <strong>TriageBench<\/strong>, a reproducible framework for evaluating LLM triage accuracy. It includes:<\/p>\n<ul>\n<li>Standard clinical dataset (Semigran vignettes)<\/li>\n<li>Paired McNemar&#8217;s test to detect model performance differences on small datasets<\/li>\n<li>Full methodology and evaluation code<\/li>\n<\/ul>\n<p>GitHub: <a href=\"https:\/\/github.com\/medaks\/medask-benchmark\">https:\/\/github.com\/medaks\/medask-benchmark<\/a><\/p>\n<p>As a demonstration, we benchmarked our own model (MedAsk) against several OpenAI models:<\/p>\n<ul>\n<li>MedAsk: <strong>87.6% accuracy<\/strong><\/li>\n<li>o3: <strong>75.6%<\/strong><\/li>\n<li>GPT\u20114.5: <strong>68.9%<\/strong><\/li>\n<\/ul>\n<p>The main limitation is dataset size (45 vignettes). We&#8217;re looking for collaborators to help expand this\u2014the field needs larger, more diverse clinical datasets.<\/p>\n<p>Blog post with full results: <a href=\"https:\/\/medask.tech\/blogs\/medical-ai-triage-accuracy-2025-medask-beats-openais-o3-gpt-4-5\/\">https:\/\/medask.tech\/blogs\/medical-ai-triage-accuracy-2025-medask-beats-openais-o3-gpt-4-5\/<\/a><\/p>\n<\/div>\n<p><!-- SC_ON -->   submitted by   <a href=\"https:\/\/www.reddit.com\/user\/Significant-Pair-275\"> \/u\/Significant-Pair-275 <\/a> <br \/> <span><a href=\"https:\/\/www.reddit.com\/r\/datasets\/comments\/1lxwioq\/we_built_an_opensource_medical_triage_benchmark\/\">[link]<\/a><\/span>   <span><a href=\"https:\/\/www.reddit.com\/r\/datasets\/comments\/1lxwioq\/we_built_an_opensource_medical_triage_benchmark\/\">[comments]<\/a><\/span><\/p><div class='watch-action'><div class='watch-position align-right'><div class='action-like'><a class='lbg-style1 like-34672 jlk' href='javascript:void(0)' data-task='like' data-post_id='34672' data-nonce='614a020375' rel='nofollow'><img class='wti-pixel' src='https:\/\/www.graviton.at\/letterswaplibrary\/wp-content\/plugins\/wti-like-post\/images\/pixel.gif' title='Like' \/><span class='lc-34672 lc'>0<\/span><\/a><\/div><\/div> <div class='status-34672 status align-right'><\/div><\/div><div class='wti-clear'><\/div>","protected":false},"excerpt":{"rendered":"<p>Medical triage means determining whether symptoms require emergency care, urgent care, or can be managed with self-care&#8230;.<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[85],"tags":[],"class_list":["post-34672","post","type-post","status-publish","format-standard","hentry","category-datatards","wpcat-85-id"],"_links":{"self":[{"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/posts\/34672","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/comments?post=34672"}],"version-history":[{"count":0,"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/posts\/34672\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/media?parent=34672"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/categories?post=34672"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/tags?post=34672"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}