A Deep Learning Pipeline for Patient Diagnosis Prediction Using - PowerPoint PPT Presentation

A Deep Learning Pipeline for Patient Diagnosis Prediction Using Electronic Health Records BIOKDD 2020 • 19th International Workshop on Data Mining in Bioinformatics Leopold Franz, Dr. Yash Raj Shrestha, Dr. Bibek Paudel 1 | | 29.05.2020 Leopold Franz

Multimorbidity: A growing problem 16% 9% 6% Ageing Population Increasing Prevalence General Increase of Multimorbidity | | 29.05.2020 2 Leopold Franz

Diagnosing Multiple Diseases Multimorbid patients are underdiagnosed 71% of the times by doctors. [Hausmann-Thürig et al., 2019] Solution: EHR Data Analytics for Diseases Diagnosis | | 29.05.2020 3 Leopold Franz

Related Work Literature Data Method Evaluation Task Structured Publication Healthcare Knowledge Transforme Readmission Mortality Diagnosis Title Authors Year Text Numerical Other FCNN RNN Other Venue Dataset graph r Prediction Prediction Detection Data Deep Patient: An Unsupervised Representation Stacked R.M., L.L., Scientific Mount Sinai to Predict the Future of Patients from the 2016 Yes Yes Denoising Yes B.K., J.D. Reports Data Warehouse Electronic Health Records [Miotto et al., 2016] Autoencoder MiME: Multilevel Medical Embedding of E.C., C.X., Med2Vec, Electronic Health Records for Predictive 2018 NeurIPS Sutter Health Yes GRU Yes W.S., J.S. GRAM Healthcare [Choi et al., 2018] Scalable and accurate deep learning with Length of A.R., E.O., npj Digital UCSF, UCM electronic health records 2018 Yes Attention TANN Stay Yes Boosted NN LSTM Yes Yes Yes K.C., J.D. Medicine Hospital Data [Rajkomar et al., 2018] Prediction Improved Hierarchical Patient Classification Length of J.K., A.R., Pre-print Hierarchical with Language Model Pretraining over Clinical Stay 2019 MIMIC-III Yes Yes Yes Yes Yes A.D. (ArXiv) RNN Notes [Kemp et al., 2019] Prediction ClinicalBERT: Modeling Clinical Notes and Language K.H., J.A., Pre-print Predicting Hospital Readmission 2019 MIMIC-III Yes Tasks (NER, Yes BERT R.R. (ArXiv) [Huang et al., 2019] RE, Q&A) | | 29.05.2020 Leopold Franz

Data-driven Methods Data Representation Learning | | 29.05.2020 5 Leopold Franz

Data MIMIC - III Intensive Care Unit (ICU) Beth Israel Deaconess Medical Center Boston, Massachusetts | | 29.05.2020 6 Leopold Franz

Representation MIMIC - III | | 29.05.2020 7 Leopold Franz

Representation | | 29.05.2020 8 Leopold Franz

Deep Learning Models Pre- 2,0.8570960175731908,0.217059 3,0.8600811744347597,0.232499 processing 4,93.94117647058823,27.890168 5,107.05833333333334,18.43740 1 6,128.0,28.160255680657446 24,1.4332023841777235,0.32984 25,152.64563907386574,56.9523 26,234.93350328315236,146.972 28,53.892805755395685,33.9506 29,-52.037351426321635,56.657 34,20.11627906976744,6.942653 Structured Data Pre- EHR processing 2 Unstructured Data | | 29.05.2020 9 Leopold Franz

Metrics Text Analytics Numerical Analytics AUPRC AUROC Recall@Prec80 | | 29.05.2020 10 Leopold Franz

Results | | 29.05.2020 11 Leopold Franz

Results | | 29.05.2020 12 Leopold Franz

Interpretation | | 29.05.2020 13 Leopold Franz

Limitations Treating Acceptance Limited Data Multimorbidity Interpretability Ethics | | 29.05.2020 14 Leopold Franz

Comments and Questions https://arxiv.org/abs/2006.16926 | | 29.05.2020 15 Leopold Franz

Appendix | | 29.05.2020 16 Leopold Franz

References [Atella et al., 2019] Atella, V., Piano Mortari, A., Kopinska, J., Belotti, F., Lapi, F., Cricelli, C., and Fontana, L. (2019). Trends in age-related disease burden and healthcare utilization. Aging cell, 18(1):e12861–e12861. 30488641[pmid]. [UN, 2020] UN (2020). World Population Ageing 2019. United Nations, Department of Economic and Social Affairs, Population Division. [WHO, 2016] WHO (2016). Multimorbidity. Technical Series on Safer Primary Care. World Health Organization. [Deetjen et al., 2020] Deetjen, U., Biesdorf, S., Guiliani, G., and Oberhänsli, W. (2020). Unleashing the power of digital health through ecosystems. [Hausmann-Thürig et al., 2019] Hausmann-Thürig, D., Kiesel, V., Zimmerli, L., Schlatter, N., von Gunten, A., Wattinger, N., and Rosemann, T. (2019). Sensitivity for multimorbid- ity: The role of diagnostic uncertainty of physicians when evaluating multimorbid video case-based vignettes. PLoS ONE, 14(4):e0215049. | | 29.05.2020 17 Leopold Franz

Literature Data Method Evaluation Task Structured Publication Healthcare Knowledge Transforme Readmission Mortality Diagnosis Title Authors Year Text Other FCNN RNN Other Numerical Venue Dataset graph r Prediction Prediction Detection Data Knowledge graph solutions in healthcare for UMLS, Build improved clinical outcomes J.A., P.M. 2018 CEUR SNOMEDCT , Yes Knowledge [Aasman and Mirhaji, 2018] OMOP Graph Unknown Learning a Health Knowledge Graph from Bayesian Build M.R, Y.H., Scientific tertiary teaching Electronic Medical Records 2017 Yes Network with Knowledge A.T., D.S. Reports hospital + [Rotmensch et al., 2017] Noisy OR gates Graph GHKG Stanford Bidirectional Summarize Learning to summarize radiology findings Y.Z., D.D., EMNLP- 2018 University Yes LSTM with Radiology [Zhang et al., 2018] T.Q., C.L. LOUHI Hospital Attention Report Deep Patient: An Unsupervised Representation Stacked R.M., L.L., Scientific Mount Sinai to Predict the Future of Patients from the Denoising 2016 Yes Yes Yes B.K., J.D. Reports Data Warehouse Electronic Health Records [Miotto et al., 2016] Autoencoder MiME: Multilevel Medical Embedding of E.C., C.X., Med2Vec, Electronic Health Records for Predictive 2018 NeurIPS Sutter Health Yes GRU Yes W.S., J.S. GRAM Healthcare [Choi et al., 2018] Neural networks versus Logistic regression for Logistic A.A., M.N., Scientific HF Dataset from 30 days all-cause readmission prediction 2019 Yes Regression, RNN, LSTM Yes G.T., M.K. Reports HCUP [Allam et al., 2019] CRF, CNN Scalable and accurate deep learning with Length of A.R., E.O., npj Digital UCSF, UCM electronic health records Stay 2018 Yes Yes Boosted NN LSTM Attention TANN Yes Yes Yes K.C., J.D. Medicine Hospital Data [Rajkomar et al., 2018] Prediction Improved Hierarchical Patient Classification Length of J.K., A.R., Pre-print Hierarchical with Language Model Pretraining over Clinical 2019 MIMIC-III Yes Yes Stay Yes Yes Yes A.D. (ArXiv) RNN Notes [Kemp et al., 2019] Prediction BioBERT: a pre-trained biomedical language PubMed Language J.L., W.Y., representation model for biomedical text mining 2019 Bioinformatics Abstracts, PMC Yes BERT Tasks (NER, S.K., J.K. [Lee et al., 2019] Articles RE, Q&A) E.A., J.M., Language Publicly Available Clinical BERT Embeddings ClinicalNLP, W.B., Tasks (NER, 2019 MIMIC-III Yes BERT [Alsentzer et al., 2019] NAACL, WS M.M. RE, Q&A) ClinicalBERT: Modeling Clinical Notes and Language K.H., J.A., Pre-print Predicting Hospital Readmission 2019 MIMIC-III Yes BERT Tasks (NER, Yes | | 29.05.2020 Leopold Franz R.R. (ArXiv) [Huang et al., 2019] RE, Q&A)

DeepObserver Pre-processing 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 5 Filter Data Time to Discharge Group Time Values Change Shape Normalize Values | | 29.05.2020 19 Leopold Franz

DeepObserver CNN Model Loss: Binary Cross Entropy Learning Rate: 10 -3 Regularization Technique Time Dimensional Filters Regularization Technique Regularization Technique Number of CCS codes to predict | | 29.05.2020 20 Leopold Franz

ClinicalBERT Pre-processing a) a) a) a) a) a) a) a) a) a) b), c) b), c) b), c) b), c) b), c) b), c) b), c) b) b) b) b) b) 1 2 3 4 5 Filter data Split data according to A) Lowercase Cut Text to equal Split train:val:test task length B) Replace abbreviations C) Remove superfluous characters | | 29.05.2020 21 Leopold Franz

ClinicalBERT_Multi Model | | 29.05.2020 22 Leopold Franz

Interpretation key query | | 29.05.2020 23 Leopold Franz