How is the Doctor Feeling? ICU Provider ... - ghassemi.xyz

1 How is the Doctor Feeling? ICU Provider Sentiment is Associated with Diagnostic Imaging Utilization Mohammad M. Ghassemi , Tuka Alhanai , Jesse D. Raffa, Roger G. Mark, Shamim Nemati and Falgun H. Chokshi Abstract The judgment of intensive care unit (ICU) Several studies over the last decade have compared the providers is difficult to measure using conventional structured diagnostic yield and cost-effectiveness of imaging utilization electronic medical record (EMR) data. However, Provider across a variety of cohort sizes (from small single center sentiment may be a proxy for such judgment. Utilizing 10.)

2 Years of EMR data, this study evaluates the association between studies [5], [6], [7], [8], [9], to large Medicare claims-mining Provider sentiment and diagnostic imaging utilization. We [10], [11]) and patient conditions (including ankle fractures extracted daily positive / negative sentiment scores of written [12], incidental lung nodules [13], altered mental status [5], Provider notes, and used a Poisson regression to estimate and head injury [14]). Most of these utilization studies are sentiment association with the total number of daily imaging performed retrospectively, using structured data from the reports.

3 After adjusting for confounding factors, we found that (1) negative sentiment was associated with increased imaging electronic medical record (EMR) or administrative claims utilization (p < ), (2) sentiment's association was most databases ( Medicare) to identify ordered examinations, pronounced at the beginning of the ICU stay (p < ), diagnoses, and sociodemographic information that are asso- and (3) the presence of any form of sentiment increased ciated with utilization rates. By relying exclusively on struc- diagnostic imaging utilization up to a critical threshold (p tured data, many retrospective analyses fail to capture the < ).

4 Our results indicate that Provider sentiment may clarify currently unexplained variance in resource utilization complete clinical context considered by health-care providers and clinical practice. when ordering imaging exams [15], [16]. At the bedside, Provider judgment reflects observations I. I NTRODUCTION that may or may not be entirely reflected in structured medi- As the United States (US) national healthcare expenditure cal data. It follows that an estimate of this judgment may help continues to rise [1], the use of high-cost medical resources explain a previously unknown component of the variance has come under increased scrutiny [2].

5 The US Health and in utilization patterns during treatment, and consequently, Human Services is already mandating more judicious use of healthcare costs. high cost medical resources, motivating investigation into the In this paper, we investigate the utility of the sentiment in historical drivers of resource utilization, and the development electronic Provider notes as a proxy of this Provider judgment of more specific criteria for justifying the utilization of high- [17]. Specifically, we investigate how Provider sentiment is cost resources [3]. One prominent example of a high cost medical resource is radiological diagnostic imaging, which accounts for nearly 10% of US health-care expenditures [4].

6 Diagnostic imaging (hereafter, imaging) is frequently cited as a high-cost medical resource in need of better defined appropriateness criteria and this year, in an effort to curtail imaging costs, the US Protecting Access to Medicare Act will mandate the use of clinical decision support tools [3]. Both authors contributed equally to this work. Ghassemi is with Department of Electrical Engineering and Com- puter Science, Massachusetts Institute of Technology,Boston, MA 02139. USA. Phone: 617-599-6010. Email: T. Alhanai is with Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology,Boston, MA 02139 USA.

7 Email: Raffa is with the Institute of Medical Engineering and Science, Massachusetts Institute of Technology,Boston, MA 02139 USA. Email: Roger G. Mark is with the Institute of Medical Engineering and Science, Massachusetts Institute of Technology,Boston, MA 02139 USA. Email: S. Nemati is with the Department of Biomedical Infor- matics, Emory University, Atlanta, GA 30322 USA. Email: Fig. 1. Sentiment in Medical Notes. Each point represents a patient's day in the ICU. Colors represent the number of radiological exams received Chokshi is with the Department of Radiology and Imaging Services, (see legend), while the size of each point indicates the number of Provider Emory University School of Medicine, Atlanta, GA 30322 USA.

8 Email: notes used to compute the sentiment. The average number of notes for each radiological exam level is shown in the figure legend. associated with diagnostic imaging utilization, after adjusting dichotomous indicators for the following conditions: obesity, for the effects of severity of illness, comorbidities, and other human immunodeficiency virus infection (HIV), metastatic factors. cancer diagnosis, diabetes and ICU type (with surgical coded Sentiment analysis is a branch of natural language pro- as one). cessing (NLP) that combines text analysis and computational linguistics to assess the emotion or polarity of a piece of text C.

9 Eligibility Criteria and Study Size (positive, negative, or neutral) [18]. Sentiment analysis has The MIMIC-III database contains notes of several distinct been used widely in non-healthcare settings, such as social categories. For this analysis we only considered Provider media [19] and newspaper publications [20] to identify and notes from the first five days of patient ICU stay which extract text-based sentiment. In the last few years this ap- were of the following types: Consult, General, Nursing, proach has also found application in the evaluation of health- Nutrition, Pharmacy, Physician, Rehabilitation Services and related topics including health and happiness [21], health Respiratory.

10 We excluded the notes of all neonatal patients care satisfaction [22], and health care reform [23]. Even and those missing any of the covariates described above. more recently, direct analysis of sentiment within medical The exclusion criteria reduced the number of distinct notes records has emerged as a topic of research. Applying word from 697,718 to 283,950, the number of distinct ICU stays embedding, Ghassemi et al. [17] explored the relationship from 52,420 to 18,607 and the number of distinct days of between Provider sentiment, patient demographics, and mor- data from 129,624 to 45,728. 76% of the excluded patients tality using sentiment analysis of structured and unstructured were neonates, while 23% were excluded due to missing EMR data of intensive care unit (ICU) patients.