RESUMEN
BACKGROUND: Triage of textual telemedical queries is a safety-critical task for medical service providers with limited remote health resources. The prioritization of patient queries containing medically severe text is necessary to optimize resource usage and provide care to those with time-sensitive needs. OBJECTIVE: We aim to evaluate the effectiveness of transfer learning solutions on the task of telemedical triage and provide a thorough error analysis, identifying telemedical queries that challenge state-of-the-art natural language processing (NLP) systems. Additionally, we aim to provide a publicly available telemedical query data set with labels for severity classification for telemedical triage of respiratory issues. METHODS: We annotated 573 medical queries from 3 online health platforms: HealthTap, HealthcareMagic, and iCliniq. We then evaluated 6 transfer learning solutions utilizing various text-embedding strategies. Specifically, we first established a baseline using a lexical classification model with term frequency-inverse document frequency (TF-IDF) features. Next, we investigated the effectiveness of global vectors for text representation (GloVe), a pretrained word-embedding method. We evaluated the performance of GloVe embeddings in the context of support vector machines (SVMs), bidirectional long short-term memory (bi-LSTM) networks, and hierarchical attention networks (HANs). Finally, we evaluated the performance of contextual text embeddings using transformer-based architectures. Specifically, we evaluated bidirectional encoder representation from transformers (BERT), Bio+Clinical-BERT, and Sentence-BERT (SBERT) on the telemedical triage task. RESULTS: We found that a simple lexical model achieved a mean F1 score of 0.865 (SD 0.048) on the telemedical triage task. GloVe-based models using SVMs, HANs, and bi-LSTMs achieved a 0.8-, 1.5-, and 2.1-point increase in the F1 score, respectively. Transformer-based models, such as BERT, Bio+Clinical-BERT, and SBERT, achieved a mean F1 score of 0.914 (SD 0.034), 0.904 (SD 0.041), and 0.917 (SD 0.037), respectively. The highest-performing model, SBERT, provided a statistically significant improvement compared to all GloVe-based and lexical baselines. However, no statistical significance was found when comparing transformer-based models. Furthermore, our error analysis revealed highly challenging query types, including those with complex negations, temporal relationships, and patient intents. CONCLUSIONS: We showed that state-of-the-art transfer learning techniques work well on the telemedical triage task, providing significant performance increase over lexical models. Additionally, we released a public telemedical triage data set using labeled questions from online medical question-and-answer (Q&A) platforms. Our analysis highlights various avenues for future works that explicitly model such query challenges.
RESUMEN
BACKGROUND: The field of dietary assessment has a long history, marked by both controversies and advances. Emerging technologies may be a potential solution to address the limitations of self-report dietary assessment methods. The Monitoring and Modeling Family Eating Dynamics (M2FED) study uses wrist-worn smartwatches to automatically detect real-time eating activity in the field. The ecological momentary assessment (EMA) methodology was also used to confirm whether eating occurred (ie, ground truth) and to measure other contextual information, including positive and negative affect, hunger, satiety, mindful eating, and social context. OBJECTIVE: This study aims to report on participant compliance (feasibility) to the 2 distinct EMA protocols of the M2FED study (hourly time-triggered and eating event-triggered assessments) and on the performance (validity) of the smartwatch algorithm in automatically detecting eating events in a family-based study. METHODS: In all, 20 families (58 participants) participated in the 2-week, observational, M2FED study. All participants wore a smartwatch on their dominant hand and responded to time-triggered and eating event-triggered mobile questionnaires via EMA while at home. Compliance to EMA was calculated overall, for hourly time-triggered mobile questionnaires, and for eating event-triggered mobile questionnaires. The predictors of compliance were determined using a logistic regression model. The number of true and false positive eating events was calculated, as well as the precision of the smartwatch algorithm. The Mann-Whitney U test, Kruskal-Wallis test, and Spearman rank correlation were used to determine whether there were differences in the detection of eating events by participant age, gender, family role, and height. RESULTS: The overall compliance rate across the 20 deployments was 89.26% (3723/4171) for all EMAs, 89.7% (3328/3710) for time-triggered EMAs, and 85.7% (395/461) for eating event-triggered EMAs. Time of day (afternoon odds ratio [OR] 0.60, 95% CI 0.42-0.85; evening OR 0.53, 95% CI 0.38-0.74) and whether other family members had also answered an EMA (OR 2.07, 95% CI 1.66-2.58) were significant predictors of compliance to time-triggered EMAs. Weekend status (OR 2.40, 95% CI 1.25-4.91) and deployment day (OR 0.92, 95% CI 0.86-0.97) were significant predictors of compliance to eating event-triggered EMAs. Participants confirmed that 76.5% (302/395) of the detected events were true eating events (ie, true positives), and the precision was 0.77. The proportion of correctly detected eating events did not significantly differ by participant age, gender, family role, or height (P>.05). CONCLUSIONS: This study demonstrates that EMA is a feasible tool to collect ground-truth eating activity and thus evaluate the performance of wearable sensors in the field. The combination of a wrist-worn smartwatch to automatically detect eating and a mobile device to capture ground-truth eating activity offers key advantages for the user and makes mobile health technologies more accessible to nonengineering behavioral researchers.