Search

The potential of artificial intelligence (AI) to reduce health care disparities and inequities is recognized, but it can also exacerbate these issues if not implemented in an equitable manner. This perspective identifies potential biases in each stage of the AI life cycle, including data collection, annotation, machine learning model development, evaluation, deployment, operationalization, monitoring, and feedback integration. To mitigate these biases, we suggest involving a diverse group of stakeholders, using human-centered AI principles. Human-centered AI can help ensure that AI systems are designed and used in a way that benefits patients and society, which can reduce health disparities and inequities. By recognizing and addressing biases at each stage of the AI life cycle, AI can achieve its potential in health care., (©You Chen, Ellen Wright Clayton, Laurie Lovett Novak, Shilo Anders, Bradley Malin. Originally published in the Journal of Medical Internet Research (https://www.jmir.org), 24.03.2023.)

Objective: There is a wide range of risk factors predisposing to the onset of neonatal encephalopathy (NE), including maternal antepartum/intrapartum comorbidities or events. However, few studies have investigated the difference in the impact of acute and chronic diseases on forecasting NE, which could assist clinicians in choosing the best course of action to prevent NE or reduce its severity and complications. In this study, we aimed to engineer features based on acute and chronic diseases and assess the differences of the impact of acute and chronic diseases on NE prediction using machine learning models., Materials and Methods: We used ten years of electronic health records of mothers from a large academic medical center to develop three types of features: chronic disease, recurrence of an acute disease, and temporal relationships between acute diseases. Two types of NE prediction models, based on acute and chronic diseases, respectively, were trained with feature selection. We further compared the prediction performance of the models with two state-of-the-art NE forecasting models. The machine learning models ranked the three types of engineered features based on their contributions to the NE prediction., Results: The NE model trained on acute disease features showed significantly higher AUC than the model relying on chronic disease features (AUC difference: 0.161, p-value < 0.001). The NE model trained on both acute and chronic disease features achieved the highest average AUC (0.889), with a significant improvement over the best existing model (0.854) with p = 0.0129. Recurrence of "known or suspected fetal abnormality affecting management of mother (655)" was assigned the highest weights in predicting NE., Conclusions: Machine learning models based on the three types of engineered features significantly improve NE prediction. Our results specifically suggest that acute disease-associated features play a more important role in predicting NE., Competing Interests: Declaration of Competing Interest The authors declare that they have no competing interests., (Copyright © 2021. Published by Elsevier B.V.)

Objective: In intensive care units (ICUs), a patient's brain function status can shift from a state of acute brain dysfunction (ABD) to one that is ABD-free and vice versa, which is challenging to forecast and, in turn, hampers the allocation of hospital resources. We aim to develop a machine learning model to predict next-day brain function status changes., Materials and Methods: Using multicenter prospective adult cohorts involving medical and surgical ICU patients from 2 civilian and 3 Veteran Affairs hospitals, we trained and externally validated a light gradient boosting machine to predict brain function status changes. We compared the performances of the boosting model against state-of-the-art models-an ABD predictive model and its variants. We applied Shapley additive explanations to identify influential factors to develop a compact model., Results: There were 1026 critically ill patients without evidence of prior major dementia, or structural brain diseases, from whom 12 295 daily transitions (ABD: 5847 days; ABD-free: 6448 days) were observed. The boosting model achieved an area under the receiver-operating characteristic curve (AUROC) of 0.824 (95% confidence interval [CI], 0.821-0.827), compared with the state-of-the-art models of 0.697 (95% CI, 0.693-0.701) with P < .001. Using 13 identified top influential factors, the compact model achieved 99.4% of the boosting model on AUROC. The boosting and the compact models demonstrated high generalizability in external validation by achieving an AUROC of 0.812 (95% CI, 0.812-0.813)., Conclusion: The inputs of the compact model are based on several simple questions that clinicians can quickly answer in practice, which demonstrates the model has direct prospective deployment potential into clinical practice, aiding in critical hospital resource allocation., (© The Author(s) 2021. Published by Oxford University Press on behalf of the American Medical Informatics Association. All rights reserved. For permissions, please email: journals.permissions@oup.com.)

Deep learning architectures have an extremely high-capacity for modeling complex data in a wide variety of domains. However, these architectures have been limited in their ability to support complex prediction problems using insurance claims data, such as readmission at 30 days, mainly due to data sparsity issue. Consequently, classical machine learning methods, especially those that embed domain knowledge in handcrafted features, are often on par with, and sometimes outperform, deep learning approaches. In this paper, we illustrate how the potential of deep learning can be achieved by blending domain knowledge within deep learning architectures to predict adverse events at hospital discharge, including readmissions. More specifically, we introduce a learning architecture that fuses a representation of patient data computed by a self-attention based recurrent neural network, with clinically relevant features. We conduct extensive experiments on a large claims dataset and show that the blended method outperforms the standard machine learning approaches., (©2021 AMIA - All rights reserved.)

Objective: Clinical corpora can be deidentified using a combination of machine-learned automated taggers and hiding in plain sight (HIPS) resynthesis. The latter replaces detected personally identifiable information (PII) with random surrogates, allowing leaked PII to blend in or "hide in plain sight." We evaluated the extent to which a malicious attacker could expose leaked PII in such a corpus., Materials and Methods: We modeled a scenario where an institution (the defender) externally shared an 800-note corpus of actual outpatient clinical encounter notes from a large, integrated health care delivery system in Washington State. These notes were deidentified by a machine-learned PII tagger and HIPS resynthesis. A malicious attacker obtained and performed a parrot attack intending to expose leaked PII in this corpus. Specifically, the attacker mimicked the defender's process by manually annotating all PII-like content in half of the released corpus, training a PII tagger on these data, and using the trained model to tag the remaining encounter notes. The attacker hypothesized that untagged identifiers would be leaked PII, discoverable by manual review. We evaluated the attacker's success using measures of leak-detection rate and accuracy., Results: The attacker correctly hypothesized that 211 (68%) of 310 actual PII leaks in the corpus were leaks, and wrongly hypothesized that 191 resynthesized PII instances were also leaks. One-third of actual leaks remained undetected., Discussion and Conclusion: A malicious parrot attack to reveal leaked PII in clinical text deidentified by machine-learned HIPS resynthesis can attenuate but not eliminate the protective effect of HIPS deidentification., (© The Author(s) 2019. Published by Oxford University Press on behalf of the American Medical Informatics Association. All rights reserved. For permissions, please email: journals.permissions@oup.com.)

Objective: User-generated content (UGC) in online environments provides opportunities to learn an individual's health status outside of clinical settings. However, the nature of UGC brings challenges in both data collecting and processing. The purpose of this study is to systematically review the effectiveness of applying machine learning (ML) methodologies to UGC for personal health investigations., Materials and Methods: We searched PubMed, Web of Science, IEEE Library, ACM library, AAAI library, and the ACL anthology. We focused on research articles that were published in English and in peer-reviewed journals or conference proceedings between 2010 and 2018. Publications that applied ML to UGC with a focus on personal health were identified for further systematic review., Results: We identified 103 eligible studies which we summarized with respect to 5 research categories, 3 data collection strategies, 3 gold standard dataset creation methods, and 4 types of features applied in ML models. Popular off-the-shelf ML models were logistic regression (n = 22), support vector machines (n = 18), naive Bayes (n = 17), ensemble learning (n = 12), and deep learning (n = 11). The most investigated problems were mental health (n = 39) and cancer (n = 15). Common health-related aspects extracted from UGC were treatment experience, sentiments and emotions, coping strategies, and social support., Conclusions: The systematic review indicated that ML can be effectively applied to UGC in facilitating the description and inference of personal health. Future research needs to focus on mitigating bias introduced when building study cohorts, creating features from free text, improving clinical creditability of UGC, and model interpretability., (© The Author(s) 2019. Published by Oxford University Press on behalf of the American Medical Informatics Association. All rights reserved. For permissions, please email: journals.permissions@oup.com.)

Objective: Data in electronic health records (EHRs) is being increasingly leveraged for secondary uses, ranging from biomedical association studies to comparative effectiveness. To perform studies at scale and transfer knowledge from one institution to another in a meaningful way, we need to harmonize the phenotypes in such systems. Traditionally, this has been accomplished through expert specification of phenotypes via standardized terminologies, such as billing codes. However, this approach may be biased by the experience and expectations of the experts, as well as the vocabulary used to describe such patients. The goal of this work is to develop a data-driven strategy to (1) infer phenotypic topics within patient populations and (2) assess the degree to which such topics facilitate a mapping across populations in disparate healthcare systems., Methods: We adapt a generative topic modeling strategy, based on latent Dirichlet allocation, to infer phenotypic topics. We utilize a variance analysis to assess the projection of a patient population from one healthcare system onto the topics learned from another system. The consistency of learned phenotypic topics was evaluated using (1) the similarity of topics, (2) the stability of a patient population across topics, and (3) the transferability of a topic across sites. We evaluated our approaches using four months of inpatient data from two geographically distinct healthcare systems: (1) Northwestern Memorial Hospital (NMH) and (2) Vanderbilt University Medical Center (VUMC)., Results: The method learned 25 phenotypic topics from each healthcare system. The average cosine similarity between matched topics across the two sites was 0.39, a remarkably high value given the very high dimensionality of the feature space. The average stability of VUMC and NMH patients across the topics of two sites was 0.988 and 0.812, respectively, as measured by the Pearson correlation coefficient. Also the VUMC and NMH topics have smaller variance of characterizing patient population of two sites than standard clinical terminologies (e.g., ICD9), suggesting they may be more reliably transferred across hospital systems., Conclusions: Phenotypic topics learned from EHR data can be more stable and transferable than billing codes for characterizing the general status of a patient population. This suggests that EHR-based research may be able to leverage such phenotypic topics as variables when pooling patient populations in predictive models., (Copyright © 2015 Elsevier Inc. All rights reserved.)

Large genomic datasets are created through numerous activities, including recreational genealogical investigations, biomedical research, and clinical care. At the same time, genomic data has become valuable for reuse beyond their initial point of collection, but privacy concerns often hinder access. Beacon services have emerged to broaden accessibility to such data. These services enable users to query for the presence of a particular minor allele in a dataset, and information helps care providers determine if genomic variation is spurious or has some known clinical indication. However, various studies have shown that this process can leak information regarding if individuals are members of the underlying dataset. There are various approaches to mitigate this vulnerability, but they are limited in that they (1) typically rely on heuristics to add noise to the Beacon responses; (2) offer probabilistic privacy guarantees only, neglecting data utility; and (3) assume a batch setting where all queries arrive at once. In this article, we present a novel algorithmic framework to ensure privacy in a Beacon service setting with a minimal number of query response flips. We represent this problem as one of combinatorial optimization in both the batch setting and the online setting (where queries arrive sequentially). We introduce principled algorithms with both privacy and, in some cases, worstcase utility guarantees. Moreover, through extensive experiments, we show that the proposed approaches significantly outperform the state of the art in terms of privacy and utility, using a dataset consisting of 800 individuals and 1.3 million single nucleotide variants. [ABSTRACT FROM AUTHOR]

Abstract: A medical specialty indicates the skills needed by health care providers to conduct key procedures or make critical judgments. However, documentation about specialties may be lacking or inaccurately specified in a health care institution. Thus, we propose to leverage diagnosis histories to recognize medical specialties that exist in practice. Such specialties that are highly recognizable through diagnosis histories are de facto diagnosis specialties. We aim to recognize de facto diagnosis specialties that are listed in the Health Care Provider Taxonomy Code Set (HPTCS) and discover those that are unlisted. First, to recognize the former, we use similarity and supervised learning models. Next, to discover de facto diagnosis specialties unlisted in the HPTCS, we introduce a general discovery‐evaluation framework. In this framework, we use a semi‐supervised learning model and an unsupervised learning model, from which the discovered specialties are subsequently evaluated by the similarity and supervised learning models used in recognition. To illustrate the potential for these approaches, we collect 2 data sets of 1 year of diagnosis histories from a large academic medical center: One is a subset of the other except for additional information useful for network analysis. The results indicate that 12 core de facto diagnosis specialties listed in the HPTCS are highly recognizable. Additionally, the semi‐supervised learning model discovers a specialty for breast cancer on the smaller data set based on network analysis, while the unsupervised learning model confirms this discovery and suggests an additional specialty for Obesity on the larger data set. The potential correctness of these 2 specialties is reinforced by the evaluation results that they are highly recognizable by similarity and supervised learning models in comparison with 12 core de facto diagnosis specialties listed in the HPTCS. [ABSTRACT FROM AUTHOR]

Cheap ubiquitous computing enables the collection of massive amounts of personal data in a wide variety of domains. Many organizations aim to share such data while obscuring features that could disclose personally identifiable information. Much of this data exhibits weak structure (e.g., text), such that machine learning approaches have been developed to detect and remove identifiers from it. While learning is never perfect, and relying on such approaches to sanitize data can leak sensitive information, a small risk is often acceptable. Our goal is to balance the value of published data and the risk of an adversary discovering leaked identifiers. We model data sanitization as a game between 1) a publisher who chooses a set of classifiers to apply to data and publishes only instances predicted as non-sensitive and 2) an attacker who combines machine learning and manual inspection to uncover leaked identifying information. We introduce a fast iterative greedy algorithm for the publisher that ensures a low utility for a resource-limited adversary. Moreover, using five text data sets we illustrate that our algorithm leaves virtually no automatically identifiable sensitive instances for a state-of-the-art learning algorithm, while sharing over 93 percent of the original data, and completes after at most five iterations. [ABSTRACT FROM AUTHOR]

Background: Biomedical research has traditionally been conducted via surveys and the analysis of medical records. However, these resources are limited in their content, such that non-traditional domains (eg, online forums and social media) have an opportunity to supplement the view of an individual's health. Objective: The objective of this study was to develop a scalable framework to detect personal health status mentions on Twitter and assess the extent to which such information is disclosed. Methods: We collected more than 250 million tweets via the Twitter streaming API over a 2-month period in 2014. The corpus was filtered down to approximately 250,000 tweets, stratified across 34 high-impact health issues, based on guidance from the Medical Expenditure Panel Survey. We created a labeled corpus of several thousand tweets via a survey, administered over Amazon Mechanical Turk, that documents when terms correspond to mentions of personal health issues or an alternative (eg, a metaphor). We engineered a scalable classifier for personal health mentions via feature selection and assessed its potential over the health issues. We further investigated the utility of the tweets by determining the extent to which Twitter users disclose personal health status. Results: Our investigation yielded several notable findings. First, we find that tweets from a small subset of the health issues can train a scalable classifier to detect health mentions. Specifically, training on 2000 tweets from four health issues (cancer, depression, hypertension, and leukemia) yielded a classifier with precision of 0.77 on all 34 health issues. Second, Twitter users disclosed personal health status for all health issues. Notably, personal health status was disclosed over 50% of the time for 11 out of 34 (33%) investigated health issues. Third, the disclosure rate was dependent on the health issue in a statistically significant manner (/P/<.001). For instance, more than 80% of the tweets about migraines (83/100) and allergies (85/100) communicated personal health status, while only around 10% of the tweets about obesity (13/100) and heart attack (12/100) did so. Fourth, the likelihood that people disclose their own versus other people's health status was dependent on health issue in a statistically significant manner as well (/P/<.001). For example, 69% (69/100) of the insomnia tweets disclosed the author's status, while only 1% (1/100) disclosed another person's status. By contrast, 1% (1/100) of the Down syndrome tweets disclosed the author's status, while 21% (21/100) disclosed another person's status. Conclusions: It is possible to automatically detect personal health status mentions on Twitter in a scalable manner. These mentions correspond to the health issues of the Twitter users themselves, but also other individuals. Though this study did not investigate the veracity of such statements, we anticipate such information may be useful in supplementing traditional health-related sources for research purposes. [ABSTRACT FROM AUTHOR]

Purpose Electronic health records contain a substantial quantity of clinical narrative, which is increasingly reused for research purposes. To share data on a large scale and respect privacy, it is critical to remove patient identifiers. De-identification tools based on machine learning have been proposed; however, model training is usually based on either a random group of documents or a pre-existing document type designation (e.g., discharge summary). This work investigates if inherent features, such as the writing complexity, can identify document subsets to enhance de-identification performance. Methods We applied an unsupervised clustering method to group two corpora based on writing complexity measures: a collection of over 4500 documents of varying document types (e.g., discharge summaries, history and physical reports, and radiology reports) from Vanderbilt University Medical Center (VUMC) and the publicly available i2b2 corpus of 889 discharge summaries. We compare the performance (via recall, precision, and F -measure) of de-identification models trained on such clusters with models trained on documents grouped randomly or VUMC document type. Results For the Vanderbilt dataset, it was observed that training and testing de-identification models on the same stylometric cluster (with the average F -measure of 0.917) tended to outperform models based on clusters of random documents (with an average F -measure of 0.881). It was further observed that increasing the size of a training subset sampled from a specific cluster could yield improved results (e.g., for subsets from a certain stylometric cluster, the F -measure raised from 0.743 to 0.841 when training size increased from 10 to 50 documents, and the F -measure reached 0.901 when the size of the training subset reached 200 documents). For the i2b2 dataset, training and testing on the same clusters based on complexity measures (average F -score 0.966) did not significantly surpass randomly selected clusters (average F -score 0.965). Conclusions Our findings illustrate that, in environments consisting of a variety of clinical documentation, de-identification models trained on writing complexity measures are better than models trained on random groups and, in many instances, document types. [ABSTRACT FROM AUTHOR]