DPM360: New Additions to Advanced Disease Progression Modeling - PowerPoint Presentation

1. DPM360: New Additions to Advanced Disease Progression ModelingAkira Koseki*, Italo Buleje*, Prithwish Chakraborty*, Elif Eyigoz*, Mohamed Ghalwash*, Takashi Itoh*, Toshiya Iwamori*, Michiharu Kudo*, Pablo Meyer*, Kenney Ng*, Parthasarathy Suryanarayanan*, Hiroki Yanagisawa*, Jianying Hu**IBM ReseacthDisease Progression Modeling (DPM) aims to characterize the progression of a disease and its comorbidities overtime using a wide range of analytics models.Typical approaches include predictive modeling, time-to-event estimation, and state-based modeling for key disease-related events.Yet we see unmet needs to facilitate the development of such advanced machine learning techniques using recent deep learning and probabilistic modeling. We explained and showed the capability of Python-based training framework for disease progression modeling using modern space-state modeling.It is believed that this can facilitate disease progression modeling using OHDSI data accumulation, especially in the Python community, and encourages the developments and commitments of advanced models leveraging opensource activities.Contact: akoseki@jp.ibm.comConclusionsBackgroundResultsMethodsFigure 1. DPM360 component view We have been developing Disease Progression Modeling Workbench 360 (DPM360) as an opensource project**. DPM360 is an easy-to-install system to help research and development of Python-based DPM models (Figure 1, FIgure2) We demonstrate advanced modeling capability including OHDISI data tooling, and extensible training framework which exploits recent achievements of state-space modeling. **https://biomedsciai.github.io/DPM360/Figure 2. General Model training pipeline of Lightsaber in DPM360 Step 1. Cohort Definition by ATLASData – MIMIC III: CCU records for 40K patientsTarget cohortOutcome cohortfirst hospital visitduration > 3 age>18Any deathStep 2. Time-series Feature Extraction in DPM360Specify OMOP Concept IDs for extraction targetRunning Jupyter NotebookDBPHeart RateBody WeightConcept IDsExtraction FeaturesStep 3. Lightsaber – versatile ML trainer in DPM360Specific Data Loader for State-Space ModelManages the training steps and tracks training statusRunning Jupyter NotebookEstimate states and transitions for patientsStep 4 DPVis – Visualizing estimated statesDetect transitions among 6 statesFound state related to death3 major states with inactive transitionsFound serious state with low BP and temperatureState descriptionState TransitionobservationsstateThick color means deviationContinuous-Time Hidden Markov Model (CTHMM) ModuleDPM360: OHDSI/OMOP Cohort definition by ATLASDPM360: Time-series Feature Extraction for OHDSI/OMOPDPM360: State-Space Model Training FrameworkDPVis: State-Space Model VisualizerCohort Definition ModulesFeature Extraction ModulesML ModulesAnd Training FrameworksPost-hoc Analysis ModulesOpen SourceProject ModulesMIMIC III DB(OMOPnized)Hidden Markov ModelFigure 2. Model training pipeline of Lightsaber in DPM360We offer and demonstrate a State-Space Modeling pipeline using recent academical achievementsA machine-learning derived Huntington Disease progression model: insights for clinical trial design”, A. Mohan, et al., Movement Disorders, 2022A probabilistic disease progression modeling approach and its application to integrated Huntington’s disease observational data”, Sun Z, et al.,  JAMIA open. 2019“Progression of type 1 diabetes from latency to symptomatic disease is predicted by distinct autoimmune trajectories”, BC Kwon et al., Nature Communications, 2022“Discovery of Parkinson’s disease states and disease progression modelling: a longitudinal data study using machine learning”, K. Severson, et al., The Lancet Digital Health, 2021