Generative transfer learning for measuring plausibility of EHR diagnosis records

Abstract

Objective Due to a complex set of processes involved with the recording of health information in the Electronic Health Records (EHRs), the truthfulness of EHR diagnosis records is questionable. We present a computational approach to estimate the probability that a single diagnosis record in the EHR reflects the true disease. Materials and Methods Using EHR data on 18 diseases from the Mass General Brigham (MGB) Biobank, we develop generative classifiers on a small set of disease-agnostic features from EHRs that aim to represent Patients, pRoviders, and their Interactions within the healthcare SysteM (PRISM features). Results We demonstrate that PRISM features and the generative PRISM classifiers are potent for estimating disease probabilities and exhibit generalizable and transferable distributional characteristics across diseases and patient populations. The joint probabilities we learn about diseases through the PRISM features via PRISM generative models are transferable and generalizable to multiple diseases. Discussion The Generative Transfer Learning (GTL) approach with PRISM classifiers enables the scalable validation of computable phenotypes in EHRs without the need for domain-specific knowledge about specific disease processes. Conclusion Probabilities computed from the generative PRISM classifier can enhance and accelerate applied Machine Learning research and discoveries with EHR data.

Evolving Phenotypes of non-hospitalized Patients that Indicate Long Covid

“…We followed the same analytic process used by Estiri et al (2021) 30 that was used to identify risk factors for COVID-19 mortality from EHR data. From the MLHO framework, the computational process to conduct multivariate PheWAS involved applying the Minimize Sparsity, Maximize Relevance (MSMR) 23,31,32 algorithm, clinical expertise, and multinomial generalized linear modeling (GLM) with component-wise functional gradient boosting, and a composite confidence score to identify the phenotypes that are positively associated with a past PCR test (see eMethods).…”

“…We followed a similar analytic process used by [31] that was used to identify risk factors for COVID-19 mortality from EHR data. From the MLHO framework, the computational process involved applying the Minimize Sparsity, Maximize Relevance (MSMR) algorithm, [23,32,33] clinical expertise, and multivariate boosting logistic regression, to compute a composite confidence score for identifying the phenotypes that are positively associated with a past RT-PCR test (see eMethods for more details).…”

“…Due to the known reliability issues of EHR diagnosis records, [33,34] we validated the phenotypes identified by MLHO through chart reviews. A clinical expert reviewed the clinical notes and longitudinal records for a random sample of five patients for each phenotype identified by MLHO with an 80-plus confidence score.…”

Machine Learning Approaches for Electronic Health Records Phenotyping: A Methodical Review

“…Among the methods used to train word embeddings, Word2vec [98] and Bidirectional Encoder Representations from Transformers (BERT) and variants [99][100][101][102] are the most frequently used (Supplementary Material Table S9). Word embeddings typically serve as the input layer to phenotyping algorithms using deep learning models and have also been used to account for ambiguous abbreviations and spelling errors in clinical notes [25,27,29,32,34,42,69,72,[74][75][76][77]83,[93][94][95][96][97][103][104][105][106][107][108][109][110][111][112][113][114].…”

“…Estiri et al utilized a self-learning approach to develop standard generative models (eg. Naive Bayes, Linear Discriminant Analysis) using a small set of labeled data (average 182 patients) and larger set of unlabeled (average 5956 patients) for classification of 18 phenotypes [114]. The approach performed on par with supervised learning, but required less labeled data (AUROC 0.78-0.99).…”