Dementia is underdiagnosed in both the general population and among Veterans. This underdiagnosis decreases quality of life, reduces opportunities for interventions, and increases health-care costs. New approaches are therefore necessary to facilitate the timely detection of dementia. This study seeks to identify cases of undiagnosed dementia by developing and validating a weakly supervised machine-learning approach that incorporates the analysis of both structured and unstructured electronic health record (EHR) data. A topic modeling approach that included latent Dirichlet allocation, stable topic extraction, and random sampling was applied to VHA EHRs. Topic features from unstructured data and features from structured data were compared between Veterans with (n = 1861) and without (n = 9305) ICD-9 dementia codes. A logistic regression model was used to develop dementia prediction scores, and manual reviews were conducted to validate the machine-learning results. A total of 853 features were identified (290 topics, 174 non-dementia ICD codes, 159 CPT codes, 59 medications, and 171 note types) for the development of logistic regression prediction scores. These scores were validated in a subset of Veterans without ICD-9 dementia codes (n = 120) by experts in dementia who performed manual record reviews and achieved a high level of inter-rater agreement. The manual reviews were used to develop a receiver of characteristic (ROC) curve with different thresholds for case detection, including a threshold of 0.061, which produced an optimal sensitivity (0.825) and specificity (0.832). Dementia is underdiagnosed, and thus, ICD codes alone cannot serve as a gold standard for diagnosis. However, this study suggests that imperfect data (e.g., ICD codes in combination with other EHR features) can serve as a silver standard to develop a risk model, apply that model to patients without dementia codes, and then select a case-detection threshold. The study is one of the first to utilize both structured and unstructured EHRs to develop risk scores for the diagnosis of dementia.

Background: Predicting the outcomes of serious mental illnesses including bipolar disorder (BD) is clinically beneficial, yet difficult. Objectives: This study aimed to predict hospitalization and mortality for patients with incident BD using a deep neural network approach. Methods: We randomly sampled 20,000 US Veterans with BD. Data on patients’ prior hospitalizations, diagnoses, procedures, medications, note types, vital signs, lab results, and BD symptoms that occurred within 1 year before and at the onset of the incident BD were extracted as features. We then created novel temporal images of patient clinical features both during the prodromal period and at the time of the disease onset. Using each temporal image as a feature, we trained and tested deep neural network learning models to predict the 1-year combined outcome of hospitalization and mortality. Results: The models achieved accuracies of 0.766–0.949 and AUCs of 0.745–0.806 for the combined outcomes. The AUC for predicting mortality was 0.814, while its highest and lowest values for predicting different types of hospitalization were 90.4% and 70.1%, suggesting that some outcomes were more difficult to predict than others. Conclusion: Deep learning using temporal graphics of clinical history is a new and promising analytical approach for mental health outcome prediction.

Transformer is the latest deep neural network (DNN) architecture for sequence data learning, which has revolutionized the field of natural language processing. This success has motivated researchers to explore its application in the healthcare domain. Despite the similarities between longitudinal clinical data and natural language data, clinical data presents unique complexities that make adapting Transformer to this domain challenging. To address this issue, we have designed a new Transformer-based DNN architecture, referred to as Hybrid Value-Aware Transformer (HVAT), which can jointly learn from longitudinal and non-longitudinal clinical data. HVAT is unique in the ability to learn from the numerical values associated with clinical codes/concepts such as labs, and in the use of a flexible longitudinal data representation called clinical tokens. We have also trained a prototype HVAT model on a case-control dataset, achieving high performance in predicting Alzheimer’s disease and related dementias as the patient outcome. The results demonstrate the potential of HVAT for broader clinical data-learning tasks.

Aims: Heart failure (HF) is a clinical syndrome with no definitive diagnostic tests. HF registries are often based on manual reviews of medical records of hospitalized HF patients identified using International Classification of Diseases (ICD) codes. However, most HF patients are not hospitalized, and manual review of big electronic health record (EHR) data is not practical. The US Department of Veterans Affairs (VA) has the largest integrated healthcare system in the nation, and an estimated 1.5 million patients have ICD codes for HF (HF ICD-code universe) in their VA EHR. The objective of our study was to develop artificial intelligence (AI) models to phenotype HF in these patients. Methods and results: The model development cohort (n = 20 000: training, 16 000; validation 2000; testing, 2000) included 10 000 patients with HF and 10 000 without HF who were matched by age, sex, race, inpatient/outpatient status, hospital, and encounter date (within 60 days). HF status was ascertained by manual chart reviews in VA's External Peer Review Program for HF (EPRP-HF) and non-HF status was ascertained by the absence of ICD codes for HF in VA EHR. Two clinicians annotated 1000 random snippets with HF-related keywords and labelled 436 as HF, which was then used to train and test a natural language processing (NLP) model to classify HF (positive predictive value or PPV, 0.81; sensitivity, 0.77). A machine learning (ML) model using linear support vector machine architecture was trained and tested to classify HF using EPRP-HF as cases (PPV, 0.86; sensitivity, 0.86). From the ‘HF ICD-code universe’, we randomly selected 200 patients (gold standard cohort) and two clinicians manually adjudicated HF (gold standard HF) in 145 of those patients by chart reviews. We calculated NLP, ML, and NLP + ML scores and used weighted F scores to derive their optimal threshold values for HF classification, which resulted in PPVs of 0.83, 0.77, and 0.85 and sensitivities of 0.86, 0.88, and 0.83, respectively. HF patients classified by the NLP + ML model were characteristically and prognostically similar to those with gold standard HF. All three models performed better than ICD code approaches: one principal hospital discharge diagnosis code for HF (PPV, 0.97; sensitivity, 0.21) or two primary outpatient encounter diagnosis codes for HF (PPV, 0.88; sensitivity, 0.54). Conclusions: These findings suggest that NLP and ML models are efficient AI tools to phenotype HF in big EHR data to create contemporary HF registries for clinical studies of effectiveness, quality improvement, and hypothesis generation.

Misspellings in clinical free text present challenges to natural language processing. With an objective to identify misspellings and their corrections, we developed a prototype spelling analysis method that implements Word2Vec, Levenshtein edit distance constraints, a lexical resource, and corpus term frequencies. We used the prototype method to process two different corpora, surgical pathology reports, and emergency department progress and visit notes, extracted from Veterans Health Administration resources. We evaluated performance by measuring positive predictive value and performing an error analysis of false positive output, using four classifications. We also performed an analysis of spelling errors in each corpus, using common error classifications. In this small-scale study utilizing a total of 76,786 clinical notes, the prototype method achieved positive predictive values of 0.9057 and 0.8979, respectively, for the surgical pathology reports, and emergency department progress and visit notes, in identifying and correcting misspelled words. False positives varied by corpus. Spelling error types were similar among the two corpora, however, the authors of emergency department progress and visit notes made over four times as many errors. Overall, the results of this study suggest that this method could also perform sufficiently in identifying misspellings in other clinical document types.