Taste

Abstract

Phenotyping electronic health records (EHR) focuses on defining meaningful patient groups (e.g., heart failure group and diabetes group) and identifying the temporal evolution of patients in those groups. Tensor factorization has been an effective tool for phenotyping. Most of the existing works assume either a static patient representation with aggregate data or only model temporal data. However, real EHR data contain both temporal (e.g., longitudinal clinical visits) and static information (e.g., patient demographics), which are difficult to model simultaneously. In this paper, we propose Temporal And Static TEnsor factorization (TASTE) that jointly models both static and temporal information to extract phenotypes. TASTE combines the PARAFAC2 model with non-negative matrix factorization to model a temporal and a static tensor. To fit the proposed model, we transform the original problem into simpler ones which are optimally solved in an alternating fashion. For each of the sub-problems, our proposed mathematical re-formulations lead to efficient sub-problem solvers. Comprehensive experiments on large EHR data from a heart failure (HF) study confirmed that TASTE is up to 14× faster than several baselines and the resulting phenotypes were confirmed to be clinically meaningful by a cardiologist. Using 60 phenotypes extracted by TASTE, a simple logistic regression can achieve the same level of area under the curve (AUC) for HF prediction compared to a deep learning model using recurrent neural networks (RNN) with 345 features.

UnsupervisedEHR‐based phenotyping via matrix and tensor decompositions

“…Sutter Palo Alto Medical Foundation Clinics; medication and diagnosis information from 50 to 80-year old adults in a heart failure study (Yin, Afshar et al, 2020); (Afshar et al, 2020); (Perros et al, 2018) CMS 3 years of claim records synthesized from 5% of the 2008 Medicare population (Ren et al, 2020;Yin, Afshar et al, 2020); (Afshar et al, 2020;Perros et al, 2018) MIMIC-II Physiologic data and vital signs time series collected from tens of thousands of ICU patient monitors (Luo et al, 2016) MIMIC-III Successor of MIMIC-II a (Joshi et al, 2016;Ren et al, 2020); (Kim, El-Kareh, et al, 2017); (Ding & Luo, 2021); (Yin, Cheung, et al, 2020); ; (Ma et al, 2019); (Yin et al, 2019); (He et al, 2019); (Yin, Afshar et al, 2020); (Zhang et al, 2021) a See https://mimic.mit.edu/docs/ for the documentation of MIMIC-II and MIMIC-III as well as for differences between the two databases. MIMIC-IV (Johnson et al, 2020), a successor of MIMIC-III, has been recently released but not yet widely used for phenotyping via low-rank approximations.…”

“…However, different-sized slabs, that is, the dimension of one mode (typically time) varies across slabs, are increasingly studied in recent years to incorporate the temporal aspect. Temporal irregularity arises, for instance, when features for subjects are recorded during multiple clinical visits, which may greatly vary across the cohort (Afshar et al, 2018(Afshar et al, , 2020Perros et al, 2017;Ren et al, 2020;Yin, Afshar et al, 2020). For regular-shaped tensors, the CP model is mostly used, while for irregular shaped tensors, which are often used for temporal phenotyping, the PARAFAC2 model is common (see Table 1).…”

“…In order to validate the phenotypes and methodology, the completed matrix is used to predict heart failure via a sparse logistic regression, and compared with other baselines (e.g., interpolation). Applying a logistic regression to the phenotype membership matrix has been used for validating phenotypes for hypertension (Henderson et al, 2017;Henderson et al, 2018), high-cost beneficiaries (Ho, Ghosh, Steinhubl, et al, 2014), mortality prediction (Luo et al, 2016;Ma et al, 2019;Yin et al, 2019;Yin, Cheung, et al, 2020) and heart failure (Afshar et al, 2020;Ho, Ghosh, Steinhubl, et al, 2014). Supervised learning as a subsequent step to an unsupervised decomposition is a common way that can indicate clinical plausibility.…”

“…However, enforcing additional sparsity can potentially make the factorization unique. Therefore, in order to provide certainty that the interpreted model and the extracted phenotypes are not arbitrary, it is crucial to discuss and investigate the uniqueness properties of matrix and tensor factorization-based approaches as in several phenotyping studies (Afshar et al, 2020;Perros et al, 2019, Perros et al, 2017Yin, Afshar et al, 2020).…”

PheME: A deep ensemble framework for improving phenotype prediction from multi-modal data

“…Automated phenotyping with machine learning provides an alternative that could be more generalized and scalable compared to rule-based algorithms [16]. Afshar et al [17] apply the tensor factorization to phenotyping tasks on structured EHRs. Suesh et al [18] use autoencoders to create lowdimensional embeddings of underlying patient phenotypes and study how different patients will react to different interventions.…”