Process Mining/ Deep Learning Approach for Healthcare Event Prediction and Occupational Safety

Details

• Personal Author:
  Pishgar M
• Description:
  With the ongoing digitization of industries and the rising number of interconnected devices, an increasing amount of system recordings are created. Process mining describes a set of techniques that help in the automatic discovery of mathematical models from such system recordings. Such system recordings could be complex including many concurrencies, noisy and infrequent behaviors. Even though many process discovery algorithms and pre-processing steps have been proposed to remove such behaviors, still none of them have been able to decrease the complexity of the process models and show the dynamic behavior of the system closely. The analysis of the discovered models might reveal important knowledge about the system behavior that is otherwise impossible to obtain. Therefore, generating a pre-processing step to improve the quality of such recording, hence improving the quality of the process model is essential. This dissertation first focuses on generating a pre-processing step, concatenation algorithm, to improve the quality of the system recordings, consequently, the process model, and improving the results of the evaluation metrics. This algorithm finds all possible combinations of concurrent events and selects several of these combination for concatenation based on a probability function, then removes all the self-loops. Significant improvements have been observed by applying concatenation algorithm on 18 benchmark datasets. Process mining techniques mainly rely on discrete event system theories in constructing the corresponding mathematical models. Therefore, they perform poorly, when continuous measures such as time and probabilities of the system events need to be considered. On the other hand, such measures can be effectively modeled and predicted through deep learning. Hence, a system that could leverage both process mining and deep learning techniques for the prediction would be useful and effective. A second contribution applies a process mining/deep learning model formalism to predict certain behaviors in healthcare systems of three real world case studies. The detailed models and the results are explained in this dissertation. For this contribution, Electronic Health Records (EHR) are first converted to the event logs. The event logs are then fed to the process discovery algorithm to produce the process model. Finally, the resultant process model, the event logs, demographics of the patients, and severity scores are fed to Decay Replay Mining (DREAM) algorithm to predict outcomes and the prediction model is evaluated through several common metrics. Significant improvements have been observed using DREAM algorithm for prediction. The third contribution of this thesis is to apply the concatenation pre-processing algorithm to real-life healthcare datasets to demonstrate its effectiveness on complex healthcare datasets. The real-life healthcare datasets are complex and noisy. Infrequent behaviors and various concurrences exist in such datasets, which causes generating inefficient and complex process model through process discovery algorithms and leading to inaccurate predictions. Therefore, it is critical to preprocess raw data to improve its quality, hence making the process model more understandable and critical predictions more accurate and trustworthy to the medical team. For this contribution, EHR are first converted to the event logs. The concatenation algorithm is then applied on the event logs. The resultant event logs are then fed to the process discovery algorithm to produce the process model. The process model is then evaluated by using common evaluation metrics. Finally, the resultant process model, the event logs, demographics of the patients, and severity scores are fed to the DREAM algorithm to predict the critical health outcomes and the prediction model is evaluated through Area Under the Curve (AUC) metric. The same procedure is tried by using raw event logs and the results are compared. Significant improvements have been observed applying concatenation as a pre-processing step and DREAM algorithms as a prediction model. The field of Artificial Intelligence (AI) is rapidly expanding, with many applications seen routinely in health care, industry, education, and increasingly in workplaces. Although there is growing evidence of applications of AI in workplaces across all industries to simplify and/or automate tasks there is a limited understanding of the role that AI contributes to address Occupational Safety and Health (OSH) concerns. Hence, a framework which reviews the application of AI in workplaces is essential to highlight the role that AI plays anticipating and controlling exposure risks in a worker's immediate environment. A fourth contribution is the creation of a framework to review the application of AI in OSH in five main industries. Risk Evolution, Detection, Evaluation, and Control of Accidents (REDECA) describes the potential applications of AI in anticipating and controlling occupational hazards and opportunities for future AI interventions. The REDECA framework has also been used to identify the existing AI solutions and specific areas where AI solutions are missed and can be developed to reduce incidents and recovery time for the agricultural tractor drivers. In this work, 260 AI papers across five sectors (oil and gas, mining, transportation, construction, and agriculture) are reviewed using the REDECA framework to highlight current applications and gaps in OSH and AI fields. As a result of this study, most of the evidence of AI in OSH research within the oil/gas and transportation sectors focuses on the development of sensors to detect hazardous situations. In construction the focus is on the use of sensors to detect incidents. The research in the agriculture sector focuses on sensors and actuators that removes workers from hazardous conditions. Application of the REDECA framework highlights AI/OSH strengths and opportunities in various industries and potential areas for collaboration. The fifth and final contribution is the case of agriculture which has the highest rates of fatality incidents in the US, 600 cases in 2019, the numbers of the tractor-related injuries have remained high with 213 cases from 1999 to 2019. Therefore, identifying root causes of agricultural tractor driver incidents, determining existing AI solutions to reduce the agricultural tractor driver incidents is needed to improve the safety of tractor drivers in the future. In this work, 171 Fatality Assessment and Control Evaluation (FACE) reports related to the tractor drivers and REDECA are used to identify the existing AI solutions and specific areas where AI solutions are missed and can be developed to reduce incidents and recovery time. Fatality reports of tractor drivers are categorized into six main categories including: run over, pinned by, fall, others (fire and crashes), roll over and overturn. Each category was then subcategorized based on similarities of incident causes in the reports. As a result, the application of the REDECA framework has revealed potential AI solutions that could improve the safety of tractor drivers. [Description provided by NIOSH]
• Subjects:
  Automation Occupational Health Safety Workplace
• Keywords:
  Algorithms Artificial Intelligence Electronic Health Records Emerging Technology Occupational Safety Risk Assessment Work Environment
• Publisher:
  University of Illinois at Chicago
• Document Type:
  Text
• Funding:
  Grant
• Genre:
  Dissertation
• Place as Subject:
  Illinois ; OSHA Region 5
• CIO:
  National Institute for Occupational Safety and Health (NIOSH)
• Topic:
  Workplace Safety & Health
• Location:
  North America
• Pages in Document:
  1-131
• NIOSHTIC Number:
  nn:20068710
• Citation:
  Chicago, IL: University of Illinois at Chicago, 2023 Apr; :1-131
• Federal Fiscal Year:
  2023
• Performing Organization:
  University of Illinois at Chicago
• Peer Reviewed:
  False
• Start Date:
  20050701
• Source Full Name:
  Process mining/ deep learning approach for healthcare event prediction and occupational safety
• End Date:
  20290630
• Collection(s):
  National Institute for Occupational Safety and Health
• Main Document Checksum:
  urn:sha-512:05db43a7ce15c08328c93a8dbf71b396995e8ce019d41826eca0cd0466ece34bff0efc10509648f9efdf88a8f88129912034616b531eead6fa32dcaa644055c9
• Download URL:
  https://stacks.cdc.gov/view/cdc/207297/cdc_207297_DS1.pdf
• File Type:
  [PDF - 1.58 MB ]