OBJECTIVE: The objective of this study is to characterize the dynamic structure of primary care consultations by identifying typical activities and their inter-relationships to inform the design of automated approaches to clinical documentation using natural language processing and summarization methods. MATERIALS AND METHODS: This is an observational study in Australian general practice involving 31 consultations with 4 primary care physicians. Consultations were audio-recorded, and computer interactions were recorded using screen capture. Physical interactions in consultation rooms were noted by observers. Brief interviews were conducted after consultations. Conversational transcripts were analyzed to identify different activities and their speech content as well as verbal cues signaling activity transitions. An activity transition analysis was then undertaken to generate a network of activities and transitions. RESULTS: Observed activity classes followed those described in well-known primary care consultation models. Activities were often fragmented across consultations, did not flow necessarily in a defined order, and the flow between activities was nonlinear. Modeling activities as a network revealed that discussing a patient's present complaint was the most central activity and was highly connected to medical history taking, physical examination, and assessment, forming a highly interrelated bundle. Family history, allergy, and investigation discussions were less connected suggesting less dependency on other activities. Clear verbal signs were often identifiable at transitions between activities. DISCUSSION: Primary care consultations do not appear to follow a classic linear model of defined information seeking activities; rather, they are fragmented, highly interdependent, and can be reactively triggered. CONCLUSION: The nonlinearity of activities has significant implications for the design of automated information capture. Whereas dictation systems generate literal translation of speech into text, speech-based clinical summary systems will need to link disparate information fragments, merge their content, and abstract coherent information summaries.

Deep learning models have achieved state-of-the-art performance in recognizing human activities, but often rely on utilizing background cues present in typical computer vision datasets that predominantly have a stationary camera. If these models are to be employed by autonomous robots in real world environments, they must be adapted to perform independently of background cues and camera motion effects. To address these challenges, we propose a new method that firstly generates generic action region proposals with good potential to locate one human action in unconstrained videos regardless of camera motion and then uses action proposals to extract and classify effective shape and motion features by a ConvNet framework. In a range of experiments, we demonstrate that by actively proposing action regions during both training and testing, state-of-the-art or better performance is achieved on benchmarks. We show the outperformance of our approach compared to the state-of-the-art in two new datasets; one emphasizes on irrelevant background, the other highlights the camera motion. We also validate our action recognition method in an abnormal behavior detection scenario to improve workplace safety. The results verify a higher success rate for our method due to the ability of our system to recognize human actions regardless of environment and camera motion.