ABSTRACT
BACKGROUND: Word embeddings are dense numeric vectors used to represent language in neural networks. Until recently, there had been no publicly released embeddings trained on clinical data. Our work is the first to study the privacy implications of releasing these models. OBJECTIVE: This paper aims to demonstrate that traditional word embeddings created on clinical corpora that have been deidentified by removing personal health information (PHI) can nonetheless be exploited to reveal sensitive patient information. METHODS: We used embeddings created from 400,000 doctor-written consultation notes and experimented with 3 common word embedding methods to explore the privacy-preserving properties of each. RESULTS: We found that if publicly released embeddings are trained from a corpus anonymized by PHI removal, it is possible to reconstruct up to 68.5% (n=411/600) of the full names that remain in the deidentified corpus and associated sensitive information to specific patients in the corpus from which the embeddings were created. We also found that the distance between the word vector representation of a patient's name and a diagnostic billing code is informative and differs significantly from the distance between the name and a code not billed for that patient. CONCLUSIONS: Special care must be taken when sharing word embeddings created from clinical texts, as current approaches may compromise patient privacy. If PHI removal is used for anonymization before traditional word embeddings are trained, it is possible to attribute sensitive information to patients who have not been fully deidentified by the (necessarily imperfect) removal algorithms. A promising alternative (ie, anonymization by PHI replacement) may avoid these flaws. Our results are timely and critical, as an increasing number of researchers are pushing for publicly available health data.
Subject(s)
Confidentiality/trends , Natural Language Processing , Algorithms , HumansABSTRACT
BACKGROUND: A verbal autopsy (VA) is a post-hoc written interview report of the symptoms preceding a person's death in cases where no official cause of death (CoD) was determined by a physician. Current leading automated VA coding methods primarily use structured data from VAs to assign a CoD category. We present a method to automatically determine CoD categories from VA free-text narratives alone. METHODS: After preprocessing and spelling correction, our method extracts word frequency counts from the narratives and uses them as input to four different machine learning classifiers: naïve Bayes, random forest, support vector machines, and a neural network. RESULTS: For individual CoD classification, our best classifier achieves a sensitivity of.770 for adult deaths for 15 CoD categories (as compared to the current best reported sensitivity of.57), and.662 with 48 WHO categories. When predicting the CoD distribution at the population level, our best classifier achieves.962 cause-specific mortality fraction accuracy for 15 categories and.908 for 48 categories, which is on par with leading CoD distribution estimation methods. CONCLUSIONS: Our narrative-based machine learning classifier performs as well as classifiers based on structured data at the individual level. Moreover, our method demonstrates that VA narratives provide important information that can be used by a machine learning system for automated CoD classification. Unlike the structured questionnaire-based methods, this method can be applied to any verbal autopsy dataset, regardless of the collection process or country of origin.
Subject(s)
Cause of Death , Diagnostic Techniques and Procedures , Machine Learning , Narration , Natural Language Processing , Adolescent , Adult , Aged , Child , Child, Preschool , Female , Humans , Infant , Infant, Newborn , Male , Middle Aged , Young AdultABSTRACT
BACKGROUND: A scoping review to characterize the literature on the use of conversations in social media as a potential source of data for detecting adverse events (AEs) related to health products. METHODS: Our specific research questions were (1) What social media listening platforms exist to detect adverse events related to health products, and what are their capabilities and characteristics? (2) What is the validity and reliability of data from social media for detecting these adverse events? MEDLINE, EMBASE, Cochrane Library, and relevant websites were searched from inception to May 2016. Any type of document (e.g., manuscripts, reports) that described the use of social media data for detecting health product AEs was included. Two reviewers independently screened citations and full-texts, and one reviewer and one verifier performed data abstraction. Descriptive synthesis was conducted. RESULTS: After screening 3631 citations and 321 full-texts, 70 unique documents with 7 companion reports available from 2001 to 2016 were included. Forty-six documents (66%) described an automated or semi-automated information extraction system to detect health product AEs from social media conversations (in the developmental phase). Seven pre-existing information extraction systems to mine social media data were identified in eight documents. Nineteen documents compared AEs reported in social media data with validated data and found consistent AE discovery in all except two documents. None of the documents reported the validity and reliability of the overall system, but some reported on the performance of individual steps in processing the data. The validity and reliability results were found for the following steps in the data processing pipeline: data de-identification (n = 1), concept identification (n = 3), concept normalization (n = 2), and relation extraction (n = 8). The methods varied widely, and some approaches yielded better results than others. CONCLUSIONS: Our results suggest that the use of social media conversations for pharmacovigilance is in its infancy. Although social media data has the potential to supplement data from regulatory agency databases; is able to capture less frequently reported AEs; and can identify AEs earlier than official alerts or regulatory changes, the utility and validity of the data source remains under-studied. TRIAL REGISTRATION: Open Science Framework ( https://osf.io/kv9hu/ ).
Subject(s)
Adverse Drug Reaction Reporting Systems , Drug-Related Side Effects and Adverse Reactions , Pharmacovigilance , Social Media , HumansABSTRACT
BACKGROUND: The recent onset of the COVID-19 pandemic and the social distancing requirement have created an increased demand for virtual support programs. Advances in artificial intelligence (AI) may offer novel solutions to management challenges such as the lack of emotional connections within virtual group interventions. Using typed text from online support groups, AI can help identify the potential risk of mental health concerns, alert group facilitator(s), and automatically recommend tailored resources while monitoring patient outcomes. OBJECTIVE: The aim of this mixed methods, single-arm study was to evaluate the feasibility, acceptability, validity, and reliability of an AI-based co-facilitator (AICF) among CancerChatCanada therapists and participants to monitor online support group participants' distress through a real-time analysis of texts posted during the support group sessions. Specifically, AICF (1) generated participant profiles with discussion topic summaries and emotion trajectories for each session, (2) identified participant(s) at risk for increased emotional distress and alerted the therapist for follow-up, and (3) automatically suggested tailored recommendations based on participant needs. Online support group participants consisted of patients with various types of cancer, and the therapists were clinically trained social workers. METHODS: Our study reports on the mixed methods evaluation of AICF, including therapists' opinions as well as quantitative measures. AICF's ability to detect distress was evaluated by the patient's real-time emoji check-in, the Linguistic Inquiry and Word Count software, and the Impact of Event Scale-Revised. RESULTS: Although quantitative results showed only some validity of AICF's ability in detecting distress, the qualitative results showed that AICF was able to detect real-time issues that are amenable to treatment, thus allowing therapists to be more proactive in supporting every group member on an individual basis. However, therapists are concerned about the ethical liability of AICF's distress detection function. CONCLUSIONS: Future works will look into wearable sensors and facial cues by using videoconferencing to overcome the barriers associated with text-based online support groups. INTERNATIONAL REGISTERED REPORT IDENTIFIER (IRRID): RR2-10.2196/21453.
ABSTRACT
BACKGROUND: The negative psychosocial impacts of cancer diagnoses and treatments are well documented. Virtual care has become an essential mode of care delivery during the COVID-19 pandemic, and online support groups (OSGs) have been shown to improve accessibility to psychosocial and supportive care. de Souza Institute offers CancerChatCanada, a therapist-led OSG service where sessions are monitored by an artificial intelligence-based co-facilitator (AICF). The AICF is equipped with a recommender system that uses natural language processing to tailor online resources to patients according to their psychosocial needs. OBJECTIVE: We aimed to outline the development protocol and evaluate the AICF on its precision and recall in recommending resources to cancer OSG members. METHODS: Human input informed the design and evaluation of the AICF on its ability to (1) appropriately identify keywords indicating a psychosocial concern and (2) recommend the most appropriate online resource to the OSG member expressing each concern. Three rounds of human evaluation and algorithm improvement were performed iteratively. RESULTS: We evaluated 7190 outputs and achieved a precision of 0.797, a recall of 0.981, and an F1 score of 0.880 by the third round of evaluation. Resources were recommended to 48 patients, and 25 (52%) accessed at least one resource. Of those who accessed the resources, 19 (75%) found them useful. CONCLUSIONS: The preliminary findings suggest that the AICF can help provide tailored support for cancer OSG members with high precision, recall, and satisfaction. The AICF has undergone rigorous human evaluation, and the results provide much-needed evidence, while outlining potential strengths and weaknesses for future applications in supportive care.
ABSTRACT
BACKGROUND: Cancer and its treatment can significantly impact the short- and long-term psychological well-being of patients and families. Emotional distress and depressive symptomatology are often associated with poor treatment adherence, reduced quality of life, and higher mortality. Cancer support groups, especially those led by health care professionals, provide a safe place for participants to discuss fear, normalize stress reactions, share solidarity, and learn about effective strategies to build resilience and enhance coping. However, in-person support groups may not always be accessible to individuals; geographic distance is one of the barriers for access, and compromised physical condition (eg, fatigue, pain) is another. Emerging evidence supports the effectiveness of online support groups in reducing access barriers. Text-based and professional-led online support groups have been offered by Cancer Chat Canada. Participants join the group discussion using text in real time. However, therapist leaders report some challenges leading text-based online support groups in the absence of visual cues, particularly in tracking participant distress. With multiple participants typing at the same time, the nuances of the text messages or red flags for distress can sometimes be missed. Recent advances in artificial intelligence such as deep learning-based natural language processing offer potential solutions. This technology can be used to analyze online support group text data to track participants' expressed emotional distress, including fear, sadness, and hopelessness. Artificial intelligence allows session activities to be monitored in real time and alerts the therapist to participant disengagement. OBJECTIVE: We aim to develop and evaluate an artificial intelligence-based cofacilitator prototype to track and monitor online support group participants' distress through real-time analysis of text-based messages posted during synchronous sessions. METHODS: An artificial intelligence-based cofacilitator will be developed to identify participants who are at-risk for increased emotional distress and track participant engagement and in-session group cohesion levels, providing real-time alerts for therapist to follow-up; generate postsession participant profiles that contain discussion content keywords and emotion profiles for each session; and automatically suggest tailored resources to participants according to their needs. The study is designed to be conducted in 4 phases consisting of (1) development based on a subset of data and an existing natural language processing framework, (2) performance evaluation using human scoring, (3) beta testing, and (4) user experience evaluation. RESULTS: This study received ethics approval in August 2019. Phase 1, development of an artificial intelligence-based cofacilitator, was completed in January 2020. As of December 2020, phase 2 is underway. The study is expected to be completed by September 2021. CONCLUSIONS: An artificial intelligence-based cofacilitator offers a promising new mode of delivery of person-centered online support groups tailored to individual needs. INTERNATIONAL REGISTERED REPORT IDENTIFIER (IRRID): DERR1-10.2196/21453.
ABSTRACT
OBJECTIVE: In this work, we introduce a privacy technique for anonymizing clinical notes that guarantees all private health information is secured (including sensitive data, such as family history, that are not adequately covered by current techniques). MATERIALS AND METHODS: We employ a new "random replacement" paradigm (replacing each token in clinical notes with neighboring word vectors from the embedding space) to achieve 100% recall on the removal of sensitive information, unachievable with current "search-and-secure" paradigms. We demonstrate the utility of this paradigm on multiple corpora in a diverse set of classification tasks. RESULTS: We empirically evaluate the effect of our anonymization technique both on upstream and downstream natural language processing tasks to show that our perturbations, while increasing security (ie, achieving 100% recall on any dataset), do not greatly impact the results of end-to-end machine learning approaches. DISCUSSION: As long as current approaches utilize precision and recall to evaluate deidentification algorithms, there will remain a risk of overlooking sensitive information. Inspired by differential privacy, we sought to make it statistically infeasible to recreate the original data, although at the cost of readability. We hope that the work will serve as a catalyst to further research into alternative deidentification methods that can address current weaknesses. CONCLUSION: Our proposed technique can secure clinical texts at a low cost and extremely high recall with a readability trade-off while remaining useful for natural language processing classification tasks. We hope that our work can be used by risk-averse data holders to release clinical texts to researchers.
Subject(s)
Algorithms , Confidentiality , Data Anonymization , Electronic Health Records , Natural Language Processing , Health Insurance Portability and Accountability Act , Health Records, Personal , Humans , United StatesABSTRACT
An impressive breadth of interdisciplinary research suggests that emotions have an influence on human behavior. Nonetheless, we still know very little about the emotional states of those actors whose daily decisions have a lasting impact on our societies: politicians in parliament. We address this question by making use of methods of natural language processing and a digitized corpus of text data spanning a century of parliamentary debates in the United Kingdom. We use this approach to examine changes in aggregate levels of emotional polarity in the British parliament, and to test a hypothesis about the emotional response of politicians to economic recessions. Our findings suggest that, contrary to popular belief, the mood of politicians has become more positive during the past decades, and that variations in emotional polarity can be predicted by the state of the national economy.
Subject(s)
Behavior , Emotions , Government Employees/psychology , Natural Language Processing , Politics , Economics , Humans , United KingdomABSTRACT
In the early stages of neurodegenerative disorders, individuals may exhibit a decline in language abilities that is difficult to quantify with standardized tests. Careful analysis of connected speech can provide valuable information about a patient's language capacities. To date, this type of analysis has been limited by its time-consuming nature. In this study, we present a method for evaluating and classifying connected speech in primary progressive aphasia using computational techniques. Syntactic and semantic features were automatically extracted from transcriptions of narrative speech for three groups: semantic dementia (SD), progressive nonfluent aphasia (PNFA), and healthy controls. Features that varied significantly between the groups were used to train machine learning classifiers, which were then tested on held-out data. We achieved accuracies well above baseline on the three binary classification tasks. An analysis of the influential features showed that in contrast with controls, both patient groups tended to use words which were higher in frequency (especially nouns for SD, and verbs for PNFA). The SD patients also tended to use words (especially nouns) that were higher in familiarity, and they produced fewer nouns, but more demonstratives and adverbs, than controls. The speech of the PNFA group tended to be slower and incorporate shorter words than controls. The patient groups were distinguished from each other by the SD patients' relatively increased use of words which are high in frequency and/or familiarity.
Subject(s)
Aphasia, Primary Progressive/diagnosis , Artificial Intelligence , Frontotemporal Dementia/diagnosis , Primary Progressive Nonfluent Aphasia/diagnosis , Semantics , Speech/physiology , Aged , Bayes Theorem , Case-Control Studies , Computer Simulation , Female , Frontotemporal Dementia/physiopathology , Humans , Language , Male , Middle Aged , Narration , Primary Progressive Nonfluent Aphasia/physiopathology , Support Vector MachineABSTRACT
In preparation for a clinical information system implementation, the Centre for Addiction and Mental Health (CAMH) Clinical Information Transformation project completed multiple preparation steps. An automated process was desired to supplement the onerous task of manual analysis of clinical forms. We used natural language processing (NLP) and machine learning (ML) methods for a series of 266 separate clinical forms. For the investigation, documents were represented by feature vectors. We used four ML algorithms for our examination of the forms: cluster analysis, k-nearest neigh-bours (kNN), decision trees and support vector machines (SVM). Parameters for each algorithm were optimized. SVM had the best performance with a precision of 64.6%. Though we did not find any method sufficiently accurate for practical use, to our knowledge this approach to forms has not been used previously in mental health.
Subject(s)
Artificial Intelligence , Documentation/classification , Forms and Records Control/methods , Mental Health/classification , Natural Language Processing , Records/classification , Vocabulary, Controlled , Humans , Pattern Recognition, Automated/methodsABSTRACT
PURPOSE: In this study, the authors explored articulatory information as a means of improving the recognition of dysarthric speech by machine. METHOD: Data were derived chiefly from the TORGO database of dysarthric articulation (Rudzicz, Namasivayam, & Wolff, 2011) in which motions of various points in the vocal tract are measured during speech. In the 1st experiment, the authors provided a baseline model indicating a relatively low performance with traditional automatic speech recognition (ASR) using only acoustic data from dysarthric individuals. In the 2nd experiment, the authors used various measures of entropy (statistical disorder) to determine whether characteristics of dysarthric articulation can reduce uncertainty in features of dysarthric acoustics. These findings led to the 3rd experiment, in which recorded dysarthric articulation was directly encoded into the speech recognition process. RESULTS: The authors found that 18.3% of the statistical disorder in the acoustics of speakers with dysarthria can be removed if articulatory parameters are known. Using articulatory models reduces phoneme recognition errors relatively by up to 6% for speakers with dysarthria in speaker-dependent systems. CONCLUSIONS: Articulatory knowledge is useful in reducing rates of error in ASR for speakers with dysarthria and in reducing statistical uncertainty of their acoustic signals. These findings may help to guide clinical decisions related to the use of ASR in the future.
Subject(s)
Cerebral Palsy/physiopathology , Dysarthria/physiopathology , Models, Biological , Speech Intelligibility/physiology , Vocal Cords/physiopathology , Adolescent , Adult , Cerebral Palsy/complications , Databases, Factual , Dysarthria/diagnosis , Dysarthria/etiology , Entropy , Female , Humans , Linear Models , Male , Markov Chains , Middle Aged , Phonetics , Speech Acoustics , Speech Production Measurement , Speech Recognition Software , Young AdultABSTRACT
Multiple pieces of text describing various pieces of evidence in clinical trials are often needed in answering a clinical question. We explore a multi-document summarization approach to automatically find this information for questions about effects of using a medication to treat a disease. Sentences in relevant documents are ranked according to various features by a machine learning approach. Those with higher scores are more important and will be included in the summary. The presence of clinical outcomes and their polarity are incorporated into the approach as features for determining importance of sentences, and the effectiveness of this is investigated, along with that of other textual features. The results show that information on clinical outcomes improves the performance of summarization.
Subject(s)
Drug Therapy , Information Storage and Retrieval/methods , Natural Language Processing , Humans , Linguistics , MEDLINEABSTRACT
Knowing the polarity of clinical outcomes is important in answering questions posed by clinicians in patient treatment. We treat analysis of this information as a classification problem. Natural language processing and machine learning techniques are applied to detect four possibilities in medical text: no outcome, positive outcome, negative outcome, and neutral outcome. A supervised learning method is used to perform the classification at the sentence level. Five feature sets are constructed: unigrams, bigrams, change phrases, negations, and categories. The performance of different combinations of feature sets is compared. The results show that generalization using the category information in the domain knowledge base Unified Medical Language System is effective in the task. The effect of context information is significant. Combining linguistic features and domain knowledge leads to the highest accuracy.
Subject(s)
Artificial Intelligence , Evidence-Based Medicine , Natural Language Processing , Humans , Linguistics , Unified Medical Language SystemABSTRACT
The high-resolution spatial induction of ultraviolet (UV) photoproducts in mammalian cellular DNA is a goal of many scientists who study UV damage and repair. Here we describe how UV photoproducts can be induced in cellular DNA within nanometre dimensions by near-diffraction-limited 750 nm infrared laser radiation. The use of multiphoton excitation to induce highly localized DNA damage in an individual cell nucleus or mitochondrion will provide much greater resolution for studies of DNA repair dynamics and intracellular localization as well as intracellular signalling processes and cell-cell communication. The technique offers an advantage over the masking method for localized irradiation of cells, as the laser radiation can specifically target a single cell and subnuclear structures such as nucleoli, nuclear membranes or any structure that can be labelled and visualized by a fluorescent tag. It also increases the time resolution with which migration of DNA repair proteins to damage sites can be monitored. We define the characteristics of localized DNA damage induction by near-infrared radiation and suggest how it may be used for new biological investigations.