ACCURACY OF A VOICE-TO-TEXT PERSONAL DICTATION SYSTEM IN THE GENERATION OF RADIOLOGY REPORTS

OBJECTIVE. Systems that convert spoken words directly into text have recently become available. The purpose of this study was to test the accuracy, on a word-for-word basis, of one such system for generating radiology reports. MATERIALS AND METHODS. The IBM Personal Dictation System (IPDS), with the optional add-on radiology vocabulary, was assessed (the system is now known as VoiceType Dictation). The system requires one to use discrete speech (i.e., with a momentary pause between words). Two hundred current radiology reports, including 100 consecutive chest radiographs, 50 consecutive thoracic CT scans, and 50 random sonography and angiography reports, were read to the system. Before testing, the IPDS had been used in the dictation of material related to chest radiology All errors were noted on a word-for-word basis and categorized as follows: incorrect, partially incorrect (no effect on meaning), not in dictionary (word was then added), homophone, or forming. Medical words (those thought to be relevant to the meaning of the report) were considered separately. Specific assessments of numbers and dates were made. Words added to the dictionary were reread after the 200 reports had been assessed. RESULTS. When all mistakes were considered, the accuracy was 0.99 for chest radiology and 0.96 for material not concerning chest radiology. When only relevant mistakes on medical words were considered, the accuracies were 0.98 and 0.96, respectively. Accuracy for numbers was 0.95 and Rr dates 0.97. Redictation of the 22 words previously not in the dictionary was 100% accurate. CONCLUSION. The IPDS is an accurate system for direct voice-to-text dictation of radiology reports that improves with continued use. The most important future enhancement of such systems will be to allow the more natural continuous or conversational speech style.