Title Page
Contents
Abstract 10
Chapter 1. Introduction 13
1. Traumatic brain injury 13
1) Clinically approaches to traumatic brain injury 13
2) Pediatric head computed tomography decision on traumatic brain injury 16
2. Artificial intelligence clinical decision support systems 22
1) A brief review of artificial intelligence clinical decision support systems 22
2) The feasibility of AI CDSS in real clinical settings 25
3. Strategy for adoption of artificial intelligence clinical decision support system in clinical settings 26
1) Quality of healthcare data and model development 26
2) Clinician's involvement for AI use in hospital settings 27
3) Future work for generalization 30
Chapter 2. Multi-task learning model to determine the need of head computed tomography for head injured patients: a nation-wide cohort study 33
1. Introduction 33
2. Methods 35
1) Datasets 35
2) Study cohort 36
3) Predictors and outcomes 37
4) Data processing 39
5) Model development 40
6) Statistical analysis 42
3. Results 42
1) Patient selection 43
2) Patient characteristics 46
3) Model performance 54
4) Comparison to the conventional clinical support tool 58
5) Model threshold selection 59
4. Discussion 62
Chapter 3. Decision effect of a deep-learning model to assist a head computed tomography order for pediatric traumatic brain injury on emergency physicians: a decision simulation study 65
1. Introduction 65
2. Methods 67
1) Participants for decision simulation study 67
2) Simulation cases selection 68
3) Process of decision simulation study 71
4) Outcomes 73
5) Survey development 73
6) Statistical method 73
3. Results 74
1) Demographics of participants 74
2) Influence of recommendation directions 75
3) Physician's factor of influence 79
4) Factors associated with AI acceptance 82
5) Survey outcome 83
4. Discussion 87
Chapter 4. International external validation and transfer learning to predict traumatic intracranial hemorrhage based on the pan-Asian trauma outcome study 91
1. Introduction 91
2. Methods 93
1) Data setting 93
2) Datasets 94
3) Variables 96
4) Outcomes 96
5) Data preprocessing 97
6) Model development and External validation 97
7) Transfer learning 101
3. Results 102
1) Population selection 102
2) Population characteristics 104
3) Model performances 110
4) Transfer learning 110
4. Conclusion 113
5. Discussion 113
Chapter 5. Conclusion 116
References 119
논문요약 128
Table 2.1. Patient's baseline characteristics 48
Table 2.2. Injury-related characteristics 51
Table 2.3. Performance using the entire study population 55
Table 2.4. Performance using subgroup (mild TBI) population 55
Table 2.5. Performance comparison with New Orleans Criteria 60
Table 2.6. Model clinical performance by age group 61
Table 3.1. Simulation case selection matrix (n＝24) 69
Table 3.2. The example of the simulation case 70
Table 3.3. Characteristics of the participants (n＝22) 75
Table 3.4. Five scale head CT ordering tendency score results on the simulation cases 78
Table 3.5. Changes in the decision to order a head CT based on the physician's characteristics 80
Table 3.6. Logistic regression analysis of factors affecting effectiveness of machine learning model to assist head CT order 82
Table 3.7. Participant's response to survey (n＝22) 85
Table 4.1. Patient's basic characteristics of merged data grouped by national registry (EDIIS) and international registry (PATOS) 105
Table 4.2. Injury related variables of merged data by EDIIS registry and PATOS registry 107
Table 4.3. multinational external validation of nationwide model (EDIIS model) 111
Table 4.4. AUROC score of each country with PATOS registry 112
Table 4.5. AUROC score of ANN Models before and after transfer learning 113
Figure 1.1. Intracerebral hemorrhage findings in head CT scan 15
Figure 1.2. Relative risk of brain tumors in relation to estimated radiation doses and brain from CT scans 17
Figure 1.3. Diagram of key interactions in knowledge-based and non-knowledge based CDSS. 24
Figure 1.4. Diagram of transfer learning 32
Figure 2.1. Model architecture 41
Figure 2.2. Patient selection diagram for two different data sources 45
Figure 2.3. Receiver operating characteristics (ROC) curve for internal and external validation outcome on the time-validation set 57
Figure 2.4. Receiver operating characteristics (ROC) curve for all outcomes on the time-validation set 57
Figure 2.5. Reliability curve (calibration plot) of the model output probabilities on the test set data 58
Figure 3.1. The example of the presentation of DEEPTICH 71
Figure 3.2. Process of simulation scenario 72
Figure 3.3. Ordering a head CT binary decision result on the simulation case 77
Figure 4.1. Process of matching variable details 99
Figure 4.2. Architecture of model validation 100
Figure 4.3. Sampling process for data selection 103