Sravan Kumar Lalam, Hari Krishna Kunderu, Shayan Ghosh, Harish Kumar A, Ashim Prasad, Francisco Lopez-Jimenez, Samir Awa… more

Electronic Health Records (EHRs) provide a rich source of medical information across different modalities such as electrocardiograms (ECG), structured EHRs (sEHR), and unstructured EHRs (text). Inspired by the fact that many cardiac and non-cardiac diseases influence the behavior of the ECG, we leverage structured EHRs and unstructured EHRs from multiple sources by pairing with ECGs and propose a set of three new multi-modal contrastive learning models that combine ECG, sEHR, and text modalities. The performance of these models is compared against different baseline models such as supervised learning models trained from scratch with random weights initialization, and self-supervised learning models trained only on ECGs. We pre-train the models on a large proprietary dataset of about 9 million ECGs from around 2.4 million patients and evaluate the pre-trained models on various downstream tasks such as classification, zero-shot retrieval, and out-of-distribution detection involving the prediction of various heart conditions using ECG waveforms as input, and demonstrate that the models presented in this work show significant improvements compared to all baseline modes.

Correspondence to: Sravan Kumar Lalam (sravankumar.l@nference.net), Melwin Babu (melwin@nference.net)

Peer-reviewed Publication: EASL SLD Summit 2023 (September 23 2023)

Preprint: EASL SLD Summit 2023 (September 23 2023)

Lars Hegstrom, Estaban Gnass, Pete Vu, Ediz Calay, Tyler Wagner, Jesse Fishman

Background and aims: The non-invasive Fibrosis-4 index (FIB-4) predicts liver fibrosis and cirrhosis in patients. FIB-4 as the primary metric for risk stratification in non-alcoholic steatohepatitis (NASH) presents several critical limitations, notably regarding the risk of cardiovascular (CV) events. Recent studies highlight the FIB-4’s potential in forecasting cardiac events, however it is critical to note that FIB-4 cutoffs as currently applied may not precisely represent the risk for these outcomes. This study aims to examine association between CV events and FIB-4 score. Method: Our study cohort comprised 1, 828 biopsy-confirmed NASH patients extracted from a deidentified, privacy-preserving database consisting of digital electronic health records (EHR) from approximately 7 million patients seen at sites over five different states from 1991 through 2020. Patients with coexisting conditions, such as hepatocellular carcinoma, alcoholism, elevated HgA1c >9% or ALT >250 U/L, viral encephalitis, bariatric surgery, and autoimmune disease, were excluded to maintain a clear focus on NASH-specific outcomes. The identification and selection of eligible patients employed a hybrid approach beginning with structured ICD diagnosis that was then enriched by natural language processing algorithms. Results: Analysis revealed patterns in the relationship between FIB-4 scores and the odds ratios of liver decompensation and CV events (myocardial infarction, congestive heart failure, unstable angina, cardiac arrest, and aneurysm dissection). For CV outcomes, the odds ratio is associated with a FIB-4 score up to 2.60, and up to 2.90 for liver decompensation. The peak FIB-4 values of 2.60 and 2.90, at which point patients are most likely to experience an undesirable outcome, are close to the guideline-recommended FIB-4 cutoff of >2.67. We observed a steep decrease in the odds ratio for cardiac outcomes at FIB-4 >2.60, while the odds ratio for liver outcomes remained steady, reinforcing a shift in the relative incidences of outcomes experienced by patients with advanced liver disease. Of note with a FIB-4 score as low as 1.0, the risk for any cardiac event is significantly elevated. This is lower than the guideline-recommended FIB-4 minimum of 1.3. Conclusion: These findings underscore the capacity of FIB-4 to forecast not only liver decompensation but also CV events in patients with NASH. The peak odds for CV events is approximately 2.60, and 2.90 for decompensation to cirrhosis, with significant risk for CV events as low as FIB-4 score of 1.0. Such findings support the need to stratify patients with an elevated FIB-4 as high-risk and also underline the critical need to revisit the lower thresholds of FIB-4 scores, now in light of considering the predictive capacity for cardiac events.

Correspondence to: Lars Hegstrom (lhegstrom@nference.net)

Peer-reviewed Publication: Transactions on Machine Learning Research (September 2023)

Uddeshya Upadhyay, Sairam Bade, Arjun Puranik, Shahir Asfahan, Melwin Babu, Francisco Lopez-Jimenez, Samuel J. Asirvatha… more

The automated analysis of medical time series, such as the electrocardiogram (ECG), electroencephalogram (EEG), pulse oximetry, etc, has the potential to serve as a valuable tool for diagnostic decisions, allowing for remote monitoring of patients and more efficient use of expensive and time-consuming medical procedures. Deep neural networks (DNNs) have been demonstrated to process such signals effectively. However, previous research has primarily focused on classifying medical time series rather than attempting to regress the continuousvalued physiological parameters central to diagnosis. One significant challenge in this regard is the imbalanced nature of the dataset, as a low prevalence of abnormal conditions can lead to heavily skewed data that results in inaccurate predictions and a lack of certainty in such predictions when deployed. To address these challenges, we propose HypUC, a framework for imbalanced probabilistic regression in medical time series, making several contributions. (i) We introduce a simple kernel density-based technique to tackle the imbalanced regression problem with medical time series. (ii) Moreover, we employ a probabilistic regression framework that allows uncertainty estimation for the predicted continuous values. (iii) We also present a new approach to calibrate the predicted uncertainty further. (iv) Finally, we demonstrate a technique to use calibrated uncertainty estimates to improve the predicted continuous value and show the efficacy of the calibrated uncertainty estimates to flag unreliable predictions. HypUC is evaluated on a large, diverse, real-world dataset of ECGs collected from millions of patients, outperforming several conventional baselines on various diagnostic tasks, suggesting potential use-case for the reliable clinical deployment of deep learning models and a prospective clinical trial. Consequently, a hyperkalemia diagnosis algorithm based on HypUC is going to be the subject of a real-world clinical prospective study.

Correspondence to: Uddeshya Upadhyay (uddeshya.upadhyay@nference.net)

Peer-reviewed Publication: American Journal of Hematology (Jul 5 2023)

Eli Silvert, Laura Hester, Eshwan Ramudu, Colin Pawlowski, Britte Kranenburg, Francis Buadi, Eli Muchtar, Samer Khaled… more

These findings demonstrate that an NLP-based approach is valuable for the comprehensive capture of signs and symptoms of AL amyloidosis from EHRs. The NLP-based method matches the quality of manual curation, but it is significantly more time-efficient and cost-effective. This analysis had several limitations. First, the lists of synonyms and ICD codes may not fully capture all terms and codes used to record the signs and symptoms. Second, this analysis only considered EHR data from a single healthcare system, and further validation studies are needed to determine if these NLP algorithms can be directly used in other healthcare systems. Going forward, an NLP method for identifying signs and symptoms from clinical notes could be integrated as part of an AL amyloidosis screening / early identification tool. These tools could reduce the time between the initial presentation of AL amyloidosis to treatment of the disease.

Correspondence to: Angela Dispenzieri (dispenzieri.angela@mayo.edu)

Peer-reviewed Publication: Journal of Clinical Oncology (June 7 2023)

Wui Ip, Colin Pawlowski, Vineet Mathew, Mayank Choudhary, Michiel Niesen, Akash Anand, Allen Mao, Ananth Peddinti, Cheta… more

Background: Unstructured EHR data contains nuanced rich insights for Real World Evidence (RWE) studies, however harnessing this Unstructured data at scale can be challenging. Here, we evaluated two Natural Language Processing-based curation (“augmented curation”) models to identify medical conditions and medications from clinical notes for patients with hepatocellular carcinoma at City of Hope National Medical Center. Methods: We deployed two augmented curation models originally trained on de-identified EHR data from the nference nSights platform, a federated data platform comprising clinical data from several Academic Medical Centers, on clinical notes found in the City of Hope POSEIDON platform, an oncology insights engine. We assessed their performances using manual chart review as reference. We then calculated an Enrichment Factor (EF) score for augmented curation model extraction and its 95% confidence intervals. The EF score was defined as the number of patients captured by either augmented curation or structured data divided by the number of patients captured by structured data alone. Results: The augmented curation models captured conditions and medications from clinical notes with F1 scores of 0.93 and 0.95, respectively. Compared to structured EHR data alone, the disease diagnosis model captured significantly more individuals with signs and symptoms such as vomiting (EF: 22.8, 95% CI: [15.2, 41.1]), weight loss (EF: 5.0, 95% CI: [4.1, 6.6]), nausea (EF: 4.3, 95% CI: [3.1, 6.9]), and edema (EF: 3.1, 95% CI: [2.7, 3.7]). Similarly, the medication use augmented curation model captured significantly more individuals with medications such as interferon, aspirin, and pembrolizumab. From a Cox proportional hazards model analysis, we found that a survival model using augmented curation-based features and structured data achieved a model concordance of 0.745 (95% CI: [0.741, 0.748]), compared to 0.722 (95% CI: [0.717, 0.726]) for the model based on structured data alone. In addition, the augmented curation-based survival model identified jaundice (HR: 2.0, 95% CI: [1.5, 2.7]) as a significant risk factor for mortality, which was not picked up by the structured data-based survival model. Conclusions: Overall, this study shows that augmented curation models can be used to accurately capture comorbidities and medications from unstructured clinical notes, and these extracted covariates are correlated with clinically meaningful endpoints such as survival. We recommend using augmented curation as a standard insight generation approach in RWE study protocols.

Correspondence to: Venky Soundararajan (venky@nference.net)