Validating a virtual human and automated feedback system for training doctor-patient communication skills

The educational intervention with the SOPHIE system has three components. The user begins by watching an instructional video about the MVP/3E’s communication paradigm followed by viewing a tutorial video on how to use the SOPHIE system. The final component of the intervention is two conversations with SOPHIE, including feedback after each conversation. SOPHIE portrays an older female patient with advanced lung cancer who is seeking information about the prognosis. The feedback page is split into 4 main sections: a transcript, and one section for each of the three E’s (see Fig. 1). The transcript section allows the user to review their conversation. Segments of the conversation where the medical professional engaged in lecturing (i.e. spoke for too long) are given a red background, and segments where the medical professional empowered the patient by asking a question are given a green background. Some segments in the transcript display suggestions for open-ended questions or empathetic statements that the medical professional could have used. The feedback system was developed through an iterative design process with close collaboration between programmers and palliative care specialists, and many of the metrics are based on statistical analysis of doctor-patient communication, as discussed in [21].

SOPHIE’s dialogue manager uses a symbolic, schema-based approach. Although LLMs have recently achieved impressive results [22][23], at the time of development, they were deemed ill-suited to this task for a variety of issues. Bender et al. argued that large language models (e.g., the current state-of-the-art) are generally insufficient for true language understanding as well as carry their own risks and potential ethical issues [24]. Large language models also come with the additional risk of going “off the rails” of the conversation parameters which makes them unpredictable and difficult to control. Without having the ability to control the dialogue, presenting the user with opportunities to practice specific communication skills poses a real challenge in application consistency. As a result of these issues, we chose to take a symbolic approach. The conversations with SOPHIE are driven by eta, which uses flexible, modifiable dialogue schema (i.e., expected event types expressed as conversational statements) to imitate natural human conversations. The dialogue manager dynamically plans and enacts the conversation in real time by combining a user interpretation process with these dialogue schema[25][26][27]. See figure 3 for an example of the dialogue and see Fig. 2 for an overview of the dialogue manager’s architecture). The user interpretation process is handled by a set of pattern transduction rules that map user utterances into simplified context-independent “gist clauses” given the immediate context of the preceding dialogue turn. The gist clause provides an explicit representation of the meaning of the user’s utterances that the system can then respond to. Response generation, which is also handled by a pattern transduction process, can involve the selection of a particular reaction by the system to the user’s gist clause, or the invocation of a new schema (e.g., the system may invoke a schema for SOPHIE discussing her medical concerns if asked a relevant question by the doctor). In the case where the system fails to extract a gist-clause, it may either ask the user to repeat and clarify their utterance, or give a generic default reaction specific to the current schema. Ultimately, a schema guided approach to dialogue management was chosen over using a large, neural language model (such as GPT3) in order to have more control over the dialogue.

Empower - We quantified empowerment using three key metrics; Questions-asked, open-questions asked and turn-taking. The need to quantify both types of questions was made apparent after consultation with Oncologists and Palliative Care specialists at University of Rochester Medical Center (URMC). A closed question helps the medical professional check patient understanding of the medical situation and prognosis while an open-ended question gives the patient an opportunity to reveal more sensitive information about their emotional state. We used expression matching to determine what type of question was asked and kept track of the total number in each category. Questions are useful for quantifying empowerment because this inevitably invites the user to take control of the conversation and thus tends to empower them. For example, by asking a question, the patient can express their concerns and voluntarily reveal external factors that would otherwise be hidden to the medical professional.

Turn-taking was quantified by keeping track of the total time for both the user and SOPHIE respectfully. Whether or not the user was lecturing for their turn was based off a previous voice study [21] and informally takes place when a medical professional is speaking for too long. This quantifies empowerment because unequal turn taking could signal a lack of empowerment for the user if their turns are too long and frequent in comparison to SOPHIE’s.

Empathize - We quantified empathize using three key metrics; sentiment trajectory, empathy word cloud and personal pronouns. Sentiment trajectory was computed based off the work of Ali et al. [21], who analyzed sentiment in the VOICE dataset led by Sen et al. [28]. They defined a “sentiment trajectory” as an average sentiment vector across time. Sentiment was computed using VADER [29] and is on a -1 to +1 scale. They used k-means clustering to identify “sytles” of sentiment trajectory and logistic regression to identify if any style was associated with good conversation outcomes where “good” is determined by the level of patient prognosis understanding. The best style was “dynamic,” having high sentiment early on in the conversation, low sentiment in the middle (likely to match patient sentiment after hearing the prognosis), and high sentiment at the end (likely to express encouragement, care and support). Our idea was that perhaps some of the complexity of empathy could be captured and quantified overall using this sentiment trajectory.

The empathy word cloud was computed using work from Sedoc et al. [30] who developed a lexicon which maps 10k words to empathy ratings on a 1-7 scale using a Mixed-Level, Feed Forward Network. We mapped every word spoken by the user through this lexicon and computed the average empathy from that. The word cloud is creating using the 15 most frequent words. We hoped that this would roughly quantify empathy for the entire conversation, but recognize the weakness in its inability to quantify statements on a sentence-level (which is where most empathy is likely to take place).

Personal pronouns were chosen to quantify empathy based of work by Sen et al. in 2017 [28] who used LIWC to analyze the correlation between different categories of words and patient ratings of doctors in the VOICE dataset. Interestingly, they found that higher rankings were correlated with the use of personal pronouns (e.g. I, You, etc.). We computed pronoun usages using NLTK part-of-speech tagging [31]. Intuitively, using pronouns makes a conversation more empathetic by appearing more personable as compared to generic, disease specific speech.

be Explicit - We quantified being Explicit using three metrics; speaking rate, reading grade and hedge words. Speaking rate was found to be associated with patient understanding [21] and is related to being explicit (e.g., if a physician speaks too slowly, or too quickly they probably are not communicating in an explicit manner). This is computed simply by taking the average words spoken per minute.

Reading grade was chosen quantify the complexity of the speech to address the issue of doctors using too much medical jargon when communicating a prognosis. Complex speech makes it more difficult for patients to understand their medical situation and we make the assumption that less complex speech is more explicit. We computed the reading grade using the standard measure of the Flesch-Kincaid readability test[32] which outputs the linguistic complexity in terms of U.S. school grades (e.g., 1st grade to 12th grade).

Hedge words were computed using a simple list of hedge words established from prior work by counting the percentage of a user’s words that appeared on the list. The most frequent 10 hedge words are made note of and a word cloud is constructed. The 2020 Horowitz et al’s paper on the MVP paradigm which SOPHIE attempts to model mentions avoiding hedging as a key part of being explicit [12]. Thus, measuring hedging helps to quantify this skill directly for the user.

We conducted an experiment consisting of 30 participants with medical backgrounds (12 medical students, 9 nurses, 4 internal medicine residents, 2 physician’s assistants, 2 psychologists, and 1 hospital chaplain). Participants were randomized (1:1 ratio) into intervention and control groups, stratified by professional/training background. The intervention group underwent the educational intervention with the SOPHIE system while the control group received no training (see Fig. 4a and 4b).

After the educational intervention, we evaluated the communication skills of the participants using human standardized patients (SPs). Every participant had a conversation with a SP via a video call. Immediately after each interaction, the SP rated the conversation using a standard scale developed with assistance from palliative care specialists. These ratings were statistically compared using a Mann Whitney U test to determine whether there were any significant differences between the the two groups as a result of undergoing the educational intervention with the SOPHIE system (see Fig. 4c). Additionally, participants in the treatment group completed a UI/UX survey designed to inform future iterations of the SOPHIE System.

SP Rating Scale - Our rating scale was developed in close collaboration with URMC Oncologists and Palliative Care Specialists. There was no existing rating scale that was appropriate as is for our experiment although prior work exists [33]. We wanted to measure whether a clinician improved based on behaviors the SOPHIE system was designed to give feedback on and reinforce. The full rating scale can be seen in table I and is based on behaviors the human SP observes during their interation.

UX/UI Rating Scale - We broke the UX scale down into three components; namely system usability, virtual human and dialogue (see Figure 5). We chose a representative sample of system usability statements from the well-established System Usability Scale (SUS)[34][35]. The statements for the virtual human and dialogue sections were developed by the research team after robust discussions. We wanted to evaluate the realism of the virtual human (e.g., the ability to look and sound like a real cancer patient). Realism in this context is meant to include holistic aspects of the interaction such as lip syncing. Discussions with our medical collaborators indicated the importance of emotional expression in real patient encounters and we therefore incorporated statements to quantify the user’s perception of the virtual human’s ability to emote. Our dialogue statements were focused on quantifying the quality of the dialogue itself. Aspects of the dialogue such as whether the responses were fluent, natural, relevant, logical and/or emotionally expressive were selected as the criteria. All statements were evaluated on a 1-5 likert scale with 5 representing strong agreement.

Our UI rating scale was also developed by the research team. We simply had the research subjects evaluate the utility of each UI element on the system feedback screen (see Fig. 1 using a 1-5 likert scale with 5 representing high utility. Likert scales allows for more nuance [36] in survey responses and are appropriate for gathering feedback for system improvement.

We used four SPs for the experiment. Each SP had an equal number of intervention and control participants ($\pm 1$). A Bonferroni corrected pairwise t-test showed no significant differences between ratings given by the different SPs.

We found that the intervention group performed significantly better on the “overall communicator” (intervention: 6.000, control: 5.067, $p<0.05$) and “aggregate score” (intervention: 36.067, control: 29.600 $p<0.05$) metrics. For every other question, there was a trend towards the intervention group, but the difference was not always statistically significant. See Table I for the full results.

Participants in the intervention group rated each feedback metric shown in figure 1 on a 1-5 Likert scale and the results can be seen in table II. Overall, we found that the most useful feedback metrics were the reading level, speaking rate, hedge words, transcript and turn-taking. The least useful metrics were positive emotion, empathy words and personal pronouns. It is important to note that user’s ratings of the feedback metrics may not equate to what they actually learned. For example, participants rated the empathy metrics relatively low, yet still performed significantly better on 2 of the 5 empathy ratings according to the Human SP evaluation.

Additionally, participants rated four components of the SOPHIE system; System Usability, Virtual Human, Dialogue, and Feedback. Every question was likewise asked on a 1-5 point Likert scale with 5 meaning most strongly agree. Fig. 5 depicts the UX experience for system usability, virtual human and dialogue. Overall, participants rated the system as easy to use with the virtual human having a realistic voice and appearance. However, the dialogue appears to be a major point of weakness in the experiment. Its responses were not rated as natural, logical, or realistic, and it did not appear to understand the user. Importantly, though, SOPHIE kept the conversation relevant, despite the variety of ways in which users could respond. SOPHIE’s ability to display emotion received mixed ratings.

Despite these limitations, debriefing interviews with participants in the intervention group indicated that participants saw overall utility in our system, and participants expressed an interest in using the tool if improvements to the dialogue and virtual human allowed the interaction to be more realistic.

Human communication is pragmatic, with patterns developing over time to become habitual and difficult to change [37]. We speculate that the extent to which an individual’s communication behaviors can be modified is dependent on how well a person’s subjective experience and recollection can be meaningfully connected to clearly-presented and actionable feedback. Based on our experiment, we see that the combination of interacting with SOPHIE and receiving automated feedback improved participant’s use of the Empower and Empathize skills and their overall communication. To what extent this increase was the result of simply interacting with SOPHIE versus receiving feedback on the interaction cannot be established based on this experiment, as we didn’t have a population which had the conversation but not the feedback. However, one indicator of the importance of feedback may be that the users rated the feedback system more highly than the dialogue and virtual human. We suspect that the feedback system is a major contributor to the differences observed. For example, a user, upon reviewing the transcript of their most recent conversation with SOPHIE and observing that they only asked three questions, may realize that they can better empower the patient by asking more questions. Similarly, as the user is reviewing the transcript, they become aware of empathetic statements they could have used. This knowledge seems to inform subsequent conversations based on the ratings from the human SPs (see q3 - asking questions - and q15 - empathetic statements - in table I). This suggests that the system’s feedback could result in an actionable plan for improvement. Further experiments with SOPHIE will be needed to confirm these intuitions about the efficacy of the feedback system.

Although we did see differences in the being Explicit skill, they were not statistically significant. This may change as we run future experiments with larger sample sizes. Additionally, we are planning further improvements to the SOPHIE system based on our UI/UX feedback from study participants (see section IV-C). Ultimately, we believe the consistency of the virtual human, dialogue and feedback system would allow a healthcare professional to hone a variety of communication skills through repeated practice.

The SOPHIE communication resources can be made accessible to anybody with a computer, microphone and internet connection. The accessibility of the system is highlighted by the fact that users rated SOPHIE as easy to use, and in particular disagreed with the statement “I needed to learn a lot of things before I could get going with this system.” The scalable, web-based nature of SOPHIE makes it a low-cost alternative to synchronous training with human SPs, and could make communication training more readily available to rural or low-income regions. This would promote equity by making communication training available to more healthcare professionals regardless of the financial resources available to them.

Specific aspects of SOPHIE could also be customized to reflect a diverse range of patients. Attributes like SOPHIE’s age, race, gender, language, and personality could be modified to represent all demographics of patient populations. Additionally, the context of the module could be readily changed. In the future, users could choose from dozens of healthcare modules focusing on specific types of conversations with customized virtual humans uniquely suited for the purpose. Different types of patient personalities could be programmed to help practice responding to different reactions and attitudes from diverse patients.

Despite efforts to improve communication skills, a clinician may fall back into their old habits unconsciously. Thus, there is a need to consistently practice these difficult conversations for maintaining, or, even enhancing, a medical professional’s skill proficiency. Future generations of SOPHIE aim to satisfy this need by improving upon the system weaknesses discovered from the UI/UX responses: namely, the dialogue management as well as the emotional expressiveness of the SOPHIE virtual human. Additionally, we plan to iterate further with medical professionals to perfect the feedback system (especially in regards to feedback elements that received low scores).

The feedback system may be extended to other applications in the future as well. For example, clinicians could have an application on their phones that could be used during real patient encounters. With the patient’s consent, a clinician could use the app to analyze the conversation. The app could generate a checklist that the clinician could quickly review to determine if they need to spend more time addressing a specific area with their patient. The data could even be tracked over time to help the clinician monitor their performance and obtain user-led, personalized insights. For example, the clinician could view the system’s recommended SOPHIE modules to refresh or improve certain communication skills. What the clinician decides to do based on this feedback is entirely their own and the design focus should always be to empower the clinician.

The limitations for this study are our small sample size, lack of clarity of which factors of the SOPHIE system caused the improvements and a lack of formal validation for the rating scales used. We elaborate on these limitations in our ethical statement which can be found after the conclusion.

Regretfully, the SOPHIE code base is not open-source. The SOPHIE project is an ongoing venture between the University of Rochester Computer Science (URCS) Department and the University of Rochester Medical Center (URMC). Once deployed, the research staff may release a starter kit for researchers who wish to run similar experiments. However, we are actively recruiting participants for clinical trials and seeking additional collaborations from other medical schools. If interested, please email Kurtis Haut at [email protected].