Evaluating large language models for lay summaries of radiology reports using tailored prompting strategies and mixed-method assessment

{"Radiology":[0],"reports":[1,53],"are":[2,18],"often":[3],"filled":[4],"with":[5,287],"medical":[6],"jargon":[7],"that":[8,183],"limits":[9],"patient":[10],"understanding.":[11],"Lay":[12],"summaries":[13,47,64,98,182,252],"can":[14],"improve":[15],"understanding":[16],"but":[17],"time-consuming":[19],"for":[20,38,178],"healthcare":[21],"providers":[22],"to":[23,31],"create.":[24],"The":[25,97],"objective":[26],"of":[27,35,144,294],"this":[28,283],"study":[29,284],"is":[30],"explore":[32],"the":[33,55,147,218,249,292],"use":[34],"tailored":[36],"prompts":[37,290],"five":[39],"Large":[40,118],"Language":[41],"Models":[42,120],"(LLMs)":[43],"in":[44,241,297],"generating":[45,179],"lay":[46,63,181,300],"from":[48,54],"radiology":[49,115],"reports.":[50],"Using":[51,141],"100":[52],"publicly":[56],"available":[57],"\\"BioNLP":[58],"2023":[59],"report":[60],"summarization\\"":[61],"dataset,":[62],"were":[65,99,151,159,170],"generated":[66,288],"by":[67,93,109,173],"each":[68],"LLM,":[69],"under":[70],"select":[71],"prompting":[72],"styles":[73],"[Few-Shot":[74],"(GPT-4),":[75],"Generated":[76],"Knowledge":[77],"(GPT-4o":[78],"mini,":[79],"Gemini":[80,84,161,224,232,244],"1.5":[81,85,162,225,233,245],"-":[82,86,124,163,166,226,234,246],"Pro,":[83],"Flash),":[87],"and":[88,117,131,137,153,155,165,176,231,271,276,291],"Zero-Shot":[89],"(Llama":[90],"3.1)]":[91],"informed":[92],"a":[94,102],"pilot":[95],"work.":[96],"evaluated":[100],"using":[101],"mixed-method":[103],"framework:":[104],"subjective":[105],"assessment":[106],"(Likert":[107],"statements)":[108],"blinded":[110],"experts":[111,175,270],"(n":[112],"=":[113],"2":[114],"fellows)":[116],"Reasoning":[119],"(LRMs)":[121],"[(Gemini":[122],"2.5":[123],"Pro":[125,167,235,247],"(LRM":[126,129,206,213,228,236,274,278],"1);":[127,193],"GPT-oss-120b":[128],"2)],":[130],"readability":[132],"metrics":[133],"(Flesch-Kincaid":[134,253],"Grade":[135,254],"Level":[136],"Flesch":[138,259],"Reading":[139,260],"Ease).":[140],"percentage":[142],"agreement":[143,266],"Likert":[145],"statements,":[146],"LLM-prompt":[148],"combinations'":[149],"performances":[150],"ranked,":[152],"Friedman":[154],"post-hoc":[156],"Nemenyi":[157],"tests":[158],"conducted.":[160],"Flash":[164,227],"(generated":[168],"knowledge)":[169],"rated":[171],"highest":[172,219],"human":[174],"LRMs":[177,272],"actionable":[180],"require":[184],"minimal":[185],"supervision":[186],"[P":[187],"<":[188,195,202,209],"4.97":[189],"×":[190,197,204,211],"10-2":[191],"(Rater":[192,199],"P":[194,201,208],"9.03":[196],"10-21":[198],"2),":[200],"6.90":[203],"10-15":[205],"1),":[207],"2.760":[210],"10-5":[212],"2).":[214],"GPT-4":[215],"(few-shot)":[216],"achieved":[217],"human-rated":[220],"accuracy":[221],"(98%),":[222],"while":[223],"1-rated:":[229],"95%)":[230],"2-rated:":[237],"91%)":[238],"ranked":[239],"first":[240],"LRM-rated":[242],"accuracy.":[243],"produced":[248],"most":[250],"accessible":[251],"Level:":[255],"7.55":[256],"±":[257,263],"1.38,":[258],"Ease:":[261],"67.84":[262],"7.78).":[264],"Strong":[265],"was":[267],"observed":[268],"between":[269],"[0.96%":[273],"1)":[275],"3.4%":[277],"2)":[279],"complete":[280],"disagreement].":[281],"Overall,":[282],"highlights":[285],"Gemini-models":[286],"knowledge":[289],"potential":[293],"LRM":[295],"evaluators":[296],"assessing":[298],"LLM-generated":[299],"summaries.":[301]}