Contextual ASR to Support EHR Adoption

描述（由申请人提供）：采用电子健康记录（EHR）系统是国家医疗保健优先事项。然而，研究表明，过渡到电子病历后，医生的生产力下降了25-40%。大多数工作流延迟是基于需要执行手动操作来填写EHR中的结构化表单，而不是传统书面笔记和转录中使用的简单非结构化叙述。先锋医疗技术公司（VMT）在美国国立卫生研究院（NIH） 1R43LM010750资助下，证明了DocTalk的可行性。DocTalk是一种实时、语音驱动、开源增强、小型实践遭遇记录系统，利用云中的集成自动语音识别（ASR）和自然语言处理（NLP），将语音从文本处理到结构化医疗数据，再到电子病历输入。虽然第一阶段的NLP准确性很高，但医生审查之前的语音准确性是不足的。幸运的是，ASR和NLP的紧密整合与医生笔记的正式结构相结合，为解决这一挑战提供了独特的基于上下文的方法。当前的语音识别方法使用单一的通用医学词典来训练识别器识别单词。医疗情境特定的概率被忽略。第一阶段SBIR项目的四个具体目标是：通过处理100万份基于文本的叙事结构化就诊记录，为患者就诊记录的每个部分创建文本语料库2。使用行业标准的开源统计语言建模工具，构建一系列特定于部分的统计语言模型（ss - slm），专门用于识别与患者就诊记录的每个特定部分相关的语音。3. 使用NLP技术从每个部分的文本中推断语言使用模式，a.检测用于调用ss - slm的触发词的部分边界b.确定每个部分的特色词分布4。评估由于使用ss - slm而提高的每个切片的准确性，目标是在相同的医疗听写系统中，与非特定切片的slm相比，总错误率减少50%。

DESCRIPTION (provided by applicant): The adoption of electronic health record (EHR) systems is a national healthcare priority. However studies show massive physician productivity drop of up to 25-40% upon transition to EHR. The majority of workflow delay is based on the need to perform manual operations to fill structured forms within the EHR, as opposed to simple unstructured narratives used in traditional written notes and transcriptions. Vanguard Medical Technologies (VMT), under NIH grant 1R43LM010750, proved feasibility for DocTalk, a real-time, speech-driven, open-source augmented, small practice encounter recording system that processes voice to text to structured medical data to EHR input, utilizing integrated automated speech recognition (ASR) and natural language processing (NLP) in the cloud. While NLP accuracy in Phase I was high, voice accuracy prior to physician review was inadequate. Fortunately, the tight integration of ASR and NLP combined with the formal structure of physician notes offers unique context based approaches to address the challenge. Current speech recognition methods use a single general-purpose medical lexicon to train a recognizer when identifying words. Medical context-specific probabilities are ignored. The four Specific Aims of this Phase I SBIR project are to: 1. Create a textual corpus for each section of a patient encounter note by processing 1 million text based narrative structured encounter notes 2. Build a family of Section-Specific Statistical Language Models (SS-SLMs) specialized in recognizing speech pertaining to each specific section of a patient encounter note, using industry standard open source statistical language modeling tools. 3. Use NLP techniques to infer patterns of language usage from text of each section, a. To detect section boundaries to be used as trigger words for invoking SS-SLMs b. To determine characteristic word distributions of each section 4. Assess improvement in accuracy per section due to use of SS-SLMs, with the goal of 50% overall reduction of errors compared to non-section-specific SLMs in the same medical dictation system. PUBLIC HEALTH RELEVANCE: Successful completion of this innovative proposed program of NLP-enhanced context based ASR, will provide the accuracy required to deploy an integrated, interactive, intuitive, low-cost data entry system for small practice primary care physicians. The augmented DocTalk system will enable physicians to increase usable information, avoid third-party transcription errors, and mitigate workflow delays. Increased small practice EHR adoption directly addresses national healthcare goals.