Influence of medical domain knowledge on deep learning for Alzheimer's disease prediction.

BACKGROUND AND OBJECTIVE: Alzheimer's disease (AD) is the most common type of dementia that can seriously affect a person's ability to perform daily activities. Estimates indicate that AD may rank third as a cause of death for older people, after heart disease and cancer. Identification of individuals at risk for developing AD is imperative for testing therapeutic interventions. The objective of the study was to determine could diagnostics of AD from EMR data alone (without relying on diagnostic imaging) be significantly improved by applying clinical domain knowledge in data preprocessing and positive dataset selection rather than setting naïve filters. METHODS: Data were extracted from the repository of heterogeneous ambulatory EMR data, collected from primary care medical offices all over the U.S. Medical domain knowledge was applied to build a positive dataset from data relevant to AD. Selected Clinically Relevant Positive (SCRP) datasets were used as inputs to a Long-Short-Term Memory (LSTM) Recurrent Neural Network (RNN) deep learning model to predict will the patient develop AD. RESULTS: Risk scores prediction of AD using the drugs domain information in an SCRP AD dataset of 2,324 patients achieved high out-of-sample score - 0.98-0.99 Area Under the Precision-Recall Curve (AUPRC) when using 90% of SCRP dataset for training. AUPRC dropped to 0.89 when training the model using less than 1,500 cases from the SCRP dataset. The model was still significantly better than when using naïve dataset selection. CONCLUSION: The LSTM RNN method that used data relevant to AD performed significantly better when learning from the SCRP dataset than when datasets were selected naïvely. The integration of qualitative medical knowledge for dataset selection and deep learning technology provided a mechanism for significant improvement of AD prediction. Accurate and early prediction of AD is significant in the identification of patients for clinical trials, which can possibly result in the discovery of new drugs for treatments of AD. Also, the contribution of the proposed predictions of AD is a better selection of patients who need imaging diagnostics for differential diagnosis of AD from other degenerative brain disorders.

Casting a Wider Net: Data Driven Discovery of Proxies for Target Diagnoses.

BACKGROUND: The Hospital Readmissions Reduction Program (HRRP) introduced in October 2012 as part of the Affordable Care Act (ACA), ties hospital reimbursement rates to adjusted 30-day readmissions and mortality performance for a small set of target diagnoses. There is growing concern and emerging evidence that use of a small set of target diagnoses to establish reimbursement rates can lead to unstable results that are susceptible to manipulation (gaming) by hospitals. METHODS: We propose a novel approach to identifying co-occurring diagnoses and procedures that can themselves serve as a proxy indicator of the target diagnosis. The proposed approach constructs a Markov Blanket that allows a high level of performance, in terms of predictive accuracy and scalability, along with interpretability of obtained results. In order to scale to a large number of co-occuring diagnoses (features) and hospital discharge records (samples), our approach begins with Google's PageRank algorithm and exploits the stability of obtained results to rank the contribution of each diagnosis/procedure in terms of presence in a Markov Blanket for outcome prediction. RESULTS: Presence of target diagnoses acute myocardial infarction (AMI), congestive heart failure (CHF), pneumonia (PN), and Sepsis in hospital discharge records for Medicare and Medicaid patients in California and New York state hospitals (2009-2011), were predicted using models trained on a subset of California state hospitals (2003-2008). Using repeated holdout evaluation, we used ~30,000,000 hospital discharge records and analyzed the stability of the proposed approach. Model performance was measured using the Area Under the ROC Curve (AUC) metric, and importance and contribution of single features to the final result. The results varied from AUC=0.68 (with SE<1e-4) for PN on cross validation datasets to AUC=0.94, with (SE<1e-7) for Sepsis on California hospitals (2009 - 2011), while the stability of features was consistently better with more training data for each target diagnosis. Prediction accuracy for considered target diagnoses approaches or exceeds accuracy estimates for discharge record data. CONCLUSIONS: This paper presents a novel approach to identifying a small subset of relevant diagnoses and procedures that approximate the Markov Blanket for target diagnoses. Accuracy and interpretability of results demonstrate the potential of our approach.