Automatically Generating Psychiatric Case Notes From Digital Transcripts of Doctor-Patient Conversations Nazmul Kazi Indika Kahanda Gianforte School of Computing Gianforte School of Computing Montana State University Montana State University Bozeman, MT, USA Bozeman, MT, USA [email protected] [email protected] Abstract Limited face-to-face time is especially disadvan- tageous for working with mental health patients Electronic health records (EHRs) are notori- where the psychiatrist could easily miss a non- ous for reducing the face-to-face time with verbal cue highly important for the correct diag- patients while increasing the screen-time for clinicians leading to burnout. This is espe- nosis. Moreover, EHR’s usability related prob- cially problematic for psychiatry care in which lems lead to unstructured and incomplete case maintaining consistent eye-contact and non- notes (Kaufman et al., 2016) which are difficult verbal cues are just as important as the spoken to search and access. words. In this ongoing work, we explore the Due to the above-mentioned downsides of feasibility of automatically generating psychi- EHRs, there have been recent attempts for de- atric EHR case notes from digital transcripts veloping novel methods for incorporating vari- of doctor-patient conversation using a two-step approach: (1) predicting semantic topics for ous techniques and technologies such as natu- segments of transcripts using supervised ma- ral language processing (NLP) for improving the chine learning, and (2) generating formal text EHR documentation process. In 2015, American of those segments using natural language pro- Medical Informatics Association reported time- cessing. Through a series of preliminary ex- consuming data entry is one of the major prob- perimental results obtained through a collec- lems in EHRs and recommended to improve EHRs tion of synthetic and real-life transcripts, we by allowing multiple modes of data entry such demonstrate the viability of this approach. as audio recording and handwritten notes (Payne 1 Introduction et al., 2015). Nagy et al.(2008) developed a voice-controlled EHR system for dentists, called An electronic health record (EHR) is a digital ver- DentVoice, that enables dentists to control the sion of a patient’s health record. EHRs were in- EHR and take notes over voice and without taking troduced as a means to improve the health care off their gloves while working with their patients. system. EHRs are real-time and store patient’s Kaufman et al.(2016) also developed an NLP- records in one place and can be shared with other enabled dictation-based data entry where clini- clinicians, researchers and authorized personals cians can write case notes over voice and able to instantly and securely. The use and implementa- reduce the time by more than 60%. tion of EHRs were spurred by the 2009 US Health Psychiatrists mostly collect information from Information Technology for Economic and Clin- their patients through conversations and these con- ical Health (HITECH) Act and 78% office-based versations are the primary source of their case clinicians reported using some form of EHR by notes. In a long-term project in collaboration with 2013 (Hsiao and Hing, 2014). National Alliance of Mental illness (NAMI) Mon- Presently, all clinicians are required to digitally tana and the Center for Mental Health Research document their interactions with their patients us- and Recovery (CMHRR) at Montana State Uni- ing EHRs. These digital documents are called case versity, we envision a pipeline that automatically notes. Manually typing case notes is time con- records a doctor-patient conversation, generates suming (Payne et al., 2015) and limits the face- the corresponding digital transcript of the conver- to-face time with their patients, which leads to sation using speech-to-text API and uses natural both patient dis-satisfaction and clinician burnout. language processing and machine learning tech- 140 Proceedings of the 2nd Clinical Natural Language Processing Workshop, pages 140–148 Minneapolis, Minnesota, June 7, 2019. c 2019 Association for Computational Linguistics niques to predict and/ or extract important pieces to generate a more formal (i.e. written) version of of information from the text. This relevant text is the text which goes in to the corresponding section then converted to a more formal written version of of the EHR form. the text and are used for auto-populating the dif- These semantic topics are suggested by the do- ferent sections of the EHR form. main experts from NAMI Montana and corre- In this work, we focus on the back-end of the spond to the main sections of a typical EHR form. above mentioned pipeline, i.e. we explore the fea- They are (1) Client details: personal information sibility of populating sections of EHR form us- of a patient, such as name, age, birth date etc., (2) ing the information extracted from a digital tran- Chief complaint: refers to the information regard- script of a doctor-patient conversation. In order ing a patient’s primary problem for which the pa- to gather gold-standard data, we develop a hu- tient is seeking medical attention., (3) Medical his- man powered digital transcript annotator and ac- tory: any past medical condition(s), treatment(s) quire annotated versions of digital transcripts of and record(s), (4) Family history: indicates medi- doctor-patient conversations with the help domain cal history of a family member of the patient, and experts. As the first step in our two-step approach, (5) Social history: refers to information about pa- we develop a machine learning model that can pre- tient’s social interactions, e.g. friends, work, fam- dict the semantic topics of segments of conversa- ily dinner etc. We call these semantic categories tions. Then we develop natural language process- “EHR categories” interchangeably. The formal ing techniques to generate a formal written text us- text is essentially the summary text that the clini- ing the corresponding segments. In this paper, we cian would write or type into the EHR form based present our preliminary findings from these two on the interaction with the patient. tasks; Figure1 depicts the high-level overview of our two-step approach. Previous studies most related to our work are (1) Lacson et al.(2006) predicting semantic topics for medical dialogue turns in the home hemodial- ysis, and (2) Wallace et al.(2014) automati- cally annotating topics in transcripts of patient- provider interactions regarding antiretroviral ad- herence. While both studies successfully use ma- chine learning for predicting semantic topics (al- beit different topics to ours) they do not focus on the development of NLP models for text summa- rization (i.e. formal text generation). The rest of the paper is structured as follows. We describe our two-step approach, data collec- tion and processing, machine learning models and natural language processing methods in chapter 2. In chapter 3, we report and discuss the perfor- mance of our methods. We summarize our find- Figure 1: High-level overview of our approach. Task1: ings, discuss limitations and potential future work Predicting EHR categories. Task 2: Formal text gen- in chapter 4. eration. ML: Machine Learning. EHR: Electronic Heallth Record. 2 Methods 2.1 Approach 2.2 Transcripts of doctor-patient dialogue As depicted in Figure1, we divide the task of Our raw dataset is composed of 18 digital tran- generating case notes from digital transcripts of scripts of doctor-patient conversations and covers doctor-patient conversations into two sub tasks: 11 presenting conditions. The presenting condi- (1) using supervised learning models to predict se- tions are Attention-deficit/ hyperactivity disorder mantic topics for segments of the transcripts and (ADHD), Alzheimer’s disease, Anger, Anorexia, then (2) using natural language processing models Anxiety, Bipolar, Borderline Personality Disor- 141 der (BPD), Depression, Obsessive Compulsive Disorder (OCD), Post Traumatic Stress Disorder (PTSD) and Schizophrenia. All transcripts are la- beled with speaker tags “Doctor:” and “Patient:” to indicate the words uttered by each individual. Thirteen of these transcripts are synthetic in that they are handwritten (i.e. typed) by a domain expert from NAMI Montana who has years of experience working with mental illness patients. Hence, each synthetic transcript represents a real case scenario of conversation between a patient (suffering from one of the presenting conditions mentioned above) and a psychiatric doctor/ clin- ician who verbally interviews the patient in a 2- person dialogue set up. Table1 reports summary statistics. Figure 2: Screen shot of the human-powered transcript annotator. Left panel displays an example transcript Rest of the five transcripts are part of Coun- while the semantic concepts are shown on the right. seling & Therapy database1 from the Alexander Street website. Hence, we refer to them as AS transcripts for the rest of the paper. Each of these the expert annotators. Any conversation pair that AS transcripts is generated from a real-life conver- was found to be irrelevant to the five categories sation between a patient and a clinician. Majority is annotated with a new category called “Others”. of these transcripts cover multiple mental condi- Conversation pair level annotations eliminated the tions. challenges in annotating a question or an answer In order to annotate transcripts using seman- on their own without the proper context provided tic topics mentioned above, we develop a human- by the preceding/ following sentences. powered transcript annotator as shown in Figure 2.3 Task 1: Predicting EHR categories 2, a responsive web application, that takes digital transcripts as input, breaks down each transcript In this task, we use the annotated digital tran- into segments where each segment starts with a scripts to generate the training data to train super- speaker tag (Doctor: or Patient:) and generates vised classification models using two different ap- samples by pairing each doctor segment with the proaches.