Deep Learning on Bone Scintigraphy to Detect Abnormal Cardiac Uptake at Risk of Cardiac Amyloidosis.

BACKGROUND: Cardiac uptake on technetium-99m whole-body scintigraphy (WBS) is almost pathognomonic of transthyretin cardiac amyloidosis. The rare false positives are often related to light-chain cardiac amyloidosis. However, this scintigraphic feature remains largely unknown, leading to misdiagnosis despite characteristic images. A retrospective review of all WBSs in a hospital database to detect those with cardiac uptake may allow the identification of undiagnosed patients. OBJECTIVES: The authors sought to develop and validate a deep learning-based model that automatically detects significant cardiac uptake (Perugini grade ≥2) on WBS from large hospital databases in order to retrieve patients at risk of cardiac amyloidosis. METHODS: The model is based on a convolutional neural network with image-level labels. The performance evaluation was performed with C-statistics using a 5-fold cross-validation scheme stratified so that the proportion of positive and negative WBSs remained constant across folds and using an external validation data set. RESULTS: The training data set consisted of 3,048 images: 281 positives (Perugini grade ≥2) and 2,767 negatives. The external validation data set consisted of 1,633 images: 102 positives and 1,531 negatives. The performance of the 5-fold cross-validation and external validation was as follows: 98.9% (± 1.0) and 96.1% for sensitivity, 99.5% (± 0.4) and 99.5% for specificity, and 0.999 (SD = 0.000) and 0.999 for the area under the curve of the receiver-operating characteristic curves. Sex, age <90 years, body mass index, injection-acquisition delay, radionuclides, and the indication of WBS only slightly affected performances. CONCLUSIONS: The authors' detection model is effective at identifying patients with cardiac uptake Perugini grade ≥2 on WBS and may help in the diagnosis of patients with cardiac amyloidosis.

The prediction of hospital length of stay using unstructured data.

OBJECTIVE: This study aimed to assess the performance improvement for machine learning-based hospital length of stay (LOS) predictions when clinical signs written in text are accounted for and compared to the traditional approach of solely considering structured information such as age, gender and major ICD diagnosis. METHODS: This study was an observational retrospective cohort study and analyzed patient stays admitted between 1 January to 24 September 2019. For each stay, a patient was admitted through the Emergency Department (ED) and stayed for more than two days in the subsequent service. LOS was predicted using two random forest models. The first included unstructured text extracted from electronic health records (EHRs). A word-embedding algorithm based on UMLS terminology with exact matching restricted to patient-centric affirmation sentences was used to assess the EHR data. The second model was primarily based on structured data in the form of diagnoses coded from the International Classification of Disease 10th Edition (ICD-10) and triage codes (CCMU/GEMSA classifications). Variables common to both models were: age, gender, zip/postal code, LOS in the ED, recent visit flag, assigned patient ward after the ED stay and short-term ED activity. Models were trained on 80% of data and performance was evaluated by accuracy on the remaining 20% test data. RESULTS: The model using unstructured data had a 75.0% accuracy compared to 74.1% for the model containing structured data. The two models produced a similar prediction in 86.6% of cases. In a secondary analysis restricted to intensive care patients, the accuracy of both models was also similar (76.3% vs 75.0%). CONCLUSIONS: LOS prediction using unstructured data had similar accuracy to using structured data and can be considered of use to accurately model LOS.