Metabolomics and a Breath Sensor Identify Acetone as a Biomarker for Heart Failure.

BACKGROUND: Multi-omics delivers more biological insight than targeted investigations. We applied multi-omics to patients with heart failure with reduced ejection fraction (HFrEF). METHODS: 46 patients with HFrEF and 20 controls underwent metabolomic profiling, including liquid/gas chromatography mass spectrometry (LC-MS/GC-MS) and solid-phase microextraction (SPME) volatilomics in plasma and urine. HFrEF was defined using left ventricular global longitudinal strain, ejection fraction and NTproBNP. A consumer breath acetone (BrACE) sensor validated results in n = 73. RESULTS: 28 metabolites were identified by GCMS, 35 by LCMS and 4 volatiles by SPME in plasma and urine. Alanine, aspartate and glutamate, citric acid cycle, arginine biosynthesis, glyoxylate and dicarboxylate metabolism were altered in HFrEF. Plasma acetone correlated with NT-proBNP (r = 0.59, 95% CI 0.4 to 0.7), 2-oxovaleric and cis-aconitic acid, involved with ketone metabolism and mitochondrial energetics. BrACE > 1.5 ppm discriminated HF from other cardiac pathology (AUC 0.8, 95% CI 0.61 to 0.92, p < 0.0001). CONCLUSION: Breath acetone discriminated HFrEF from other cardiac pathology using a consumer sensor, but was not cardiac specific.

A machine learning model for predicting risk of hospital readmission within 30 days of discharge: validated with LACE index and patient at risk of hospital readmission (PARR) model.

The objective of this study was to design and develop a predictive model for 30-day risk of hospital readmission using machine learning techniques. The proposed predictive model was then validated with the two most commonly used risk of readmission models: LACE index and patient at risk of hospital readmission (PARR). The study cohort consisted of 180,118 admissions with 22,565 (12.5%) of actual readmissions within 30 days of hospital discharge, from 01 Jan 2015 to 31 Dec 2016 from two Auckland-region hospitals. We developed a machine learning model to predict 30-day readmissions using the model types XGBoost, Random Forests, and Adaboost with decision stumps as a base learner with different feature combinations and preprocessing procedures. The proposed model achieved the F1-score (0.386 ± 0.006), sensitivity (0.598 ± 0.013), positive predictive value (PPV) (0.285 ± 0.004), and negative predictive value (NPV) (0.932 ± 0.002). When compared with LACE and PARR(NZ) models, the proposed model achieved better F1-score by 12.7% compared with LACE and 23.2% compared with PARR(NZ). The mean sensitivity of the proposed model was 6.0% higher than LACE and 41% higher than PARR(NZ). The mean PPV was 15.9% and 14.6% higher than LACE and PARR(NZ) respectively. We presented an all-cause predictive model for 30-day risk of hospital readmission with an area under the receiver operating characteristics (AUROC) of 0.75 for the entire dataset. Graphical abstract.

Predicting Patients at Risk of 30-Day Unplanned Hospital Readmission.

We developed a machine learning model to predict 30-day readmissions using the model types; XGBoost, Random Forests and Adaboost with decision stumps as a base learner with different feature combinations and preprocessing procedures. The proposed model achieved the F1-score (0.386 ± 0.006), sensitivity (0.598 ± 0.013), positive predictive value (PPV) (0.285 ± 0.004) and negative predictive value (NPV) (0.932 ± 0.002). When compared with LACE and PARR (NZ) models, the proposed model achieved better F1-score by 12.5% compared to LACE and 22.9% compared to PARR (NZ). The mean sensitivity of the proposed model was 6.0% higher than LACE and 42.4% higher than PARR (NZ). The mean PPV was 15.9% and 13.5% higher than LACE and PARR (NZ) respectively.

Machine Learning-based Risk of Hospital Readmissions: Predicting Acute Readmissions within 30 Days of Discharge.

The objective of this study was to design and develop a 30-day risk of hospital readmission predictive model using machine learning techniques. The proposed risk of readmission predictive model was then validated with the two most commonly used risk of readmission models - LACE index and patient at-risk of hospital readmission (PARR). The study cohort consisted of 180,118 admissions with 22565 (12.5%) of actual readmissions within 30-day of hospital discharge, from 01 Jan 2015 to 31 Dec 2016 from two Auckland-region hospitals. We developed a machine learning model to predict 30-day readmissions using the model types: XGBoost, Random Forests and Adaboost with decision stumps as a base learner with different feature combinations and preprocessing procedures. The proposed model achieved the F1-score (0.386 ± 0.006), sensitivity (0.598 ± 0.013), positive predictive value (PPV) (0.285 ± 0.004) and negative predictive value (NPV) (0.932 ± 0.002). When compared with LACE and PARR (NZ) models, the proposed model achieved better F1-score by 12.5% compared to LACE and 22.9% compared to PARR (NZ). The mean sensitivity of the proposed model was 6.0% higher than LACE and 42.4% higher than PARR (NZ). The mean PPV was 15.9% and 13.5% higher than LACE and PARR (NZ) respectively.

Evaluation of Patients at Risk of Hospital Readmission (PARR) and LACE Risk Score for New Zealand Context.

Identification and prediction of patients who are at risk of hospital readmission is a critical step towards the reduction of the potential avoidable costs for healthcare organisations. This research was focused on the evaluation of LACE Index for Readmission - Length of stay (days), Acute (emergent) admission, Charlson Comorbidity Index and number of ED visits within six months (LACE) and Patients At Risk of Hospital Readmission (PARR) using New Zealand hospital admissions. This research estimates the risk for all readmissions rather than only those in a subset of referenced conditions. In total, 213,440 admissions between 1 Jan 2015 and 31 Dec 2016 were selected after appropriate ethics approvals and permissions from the three hospitals. The evaluation method used is the Receiver Operating Characteristics (ROC) analysis to evaluate the accuracy of both the LACE and PARR models. As a result, The LACE index achieved an Area Under the Curve (AUC) score of 0.658 in predicting 30-day readmissions. The optimal cut-off for the LACE index is a score of 7 or more with sensitivity of 0.752 and specificity of 0.564. Whereas, the PARR algorithm achieved an AUC score of 0.628 in predicting 30-day readmissions. The optimal cut-off for the PARR index is a score of 0.34 or more with sensitivity of 0.714 and specificity of 0.542.

Improving Clinical Named-Entity Recognition with Transfer Learning.

Transfer learning is a powerful machine learning technique that enables the internalizing and reuse of prior knowledge to new tasks. Transfer learning is currently the starting point for recognition tasks such as computer vision. However, in natural language processing (NLP), the application of this technique is less prevalent. Our research investigates how, through the application of transfer learning, existing knowledge can be used to build more accurate NLP models. We subsequently applied these models to a named-entity recognition (NER) task. Our experimental results show significantly better recognition performance can be obtained through leveraging knowledge from a base model, trained using poorly annotated data.