2022 Fiscal Year Research-status Report
Development of a novel method for prediction using artificial image and image identification
| Project/Area Number |
22K21186
|
|---|
| Research Institution | Fujita Health University |
|---|
Principal Investigator |
HE YUPENG 藤田医科大学, 医学部, 助教 (00953267)
|
|---|
| Project Period (FY) |
2022-08-31 – 2024-03-31
|
| Keywords | Artifical image / Model performance / Prediction / Epidemiology |
|---|
| Outline of Annual Research Achievements |
In our current research, we have discovered the possibility of creating the optimal artificial image. Our research plan involved converting each pixel from a corresponding feature variable and using all these variables (a total of n) to construct artificial images. However, the vast number of possibilities (n factorial possibilities) made it infeasible to iterate through all of them with current computing abilities. Thereafter, we transformed this two-dimensional problem of finding the optimal artificial image into a one-dimensional problem that proves the possibility of the optimal artificial image.
Our findings demonstrate that changing the order of the feature variables when constructing the model affects the prediction accuracy, and this accuracy follows a normal distribution. Therefore, we believe that the feature order resulting in the highest prediction accuracy can be used to construct the optimal artificial image. This research provides important insights for machine learning and epidemiological research utilizing machine learning tools. In particular, as the number of independent variables increases significantly, the order becomes a crucial factor for improving model performance.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
We are sincerely grateful to KAKENHI for funding this research, as it has greatly contributed to the smooth progress of our study. Overall, the experimental process has been proceeding systematically according to the original plan. However, there have been some modifications to the plan, particularly in terms of the selection of the experimental dataset. To be more specific, we have chosen a new dataset focusing on the physical, psychiatric, and social comorbidities of individuals with schizophrenia in Japan. This new dataset offers several advantages over the original one. Firstly, it contains a larger number of features, which enhances the interpretability of the generated artificial images. Secondly, the features are more closely associated with the response, leading to greater stability during model training and testing. Furthermore, the dataset as a whole is cleaner and devoid of distracting elements such as missing values, reducing the time required for data cleaning. Additionally, the utilization of current generative AI technology has proven beneficial, particularly in terms of programming and code error detection throughout the experiment. However, it's important to note that we have encountered an unexpected challenge: the computational workload is greater than initially anticipated. In the upcoming stages of the experiment, we will implement appropriate measures to address this issue effectively.
|
| Strategy for Future Research Activity |
According to the plan, our future research will focus on several key aspects, including identifying the information order (feature order), visualizing images, and assessing the practicality of this novel method. While our current findings have confirmed the impact of feature order on model training and testing at the experimental level, we aim to seek additional theoretical support from mathematics or statistics in the next phase of our research. Additionally, considering the limitations in computing power, we will explore the utilization of alternative algorithms (such as genetic algorithms) to establish the optimal feature order and construct artificial images based on it.
In terms of application, we will strive to elucidate the relationship between the features and response through the analysis of the established artificial images. This represents a crucial step in applying our novel method to disease understanding and future risk prediction. Due to time constraints in the early stages, we were unable to release our findings through journals or conferences. However, in our subsequent research, we will expedite the publication process and actively seek more peer reviews and valuable suggestions through academic exchanges.
|
| Causes of Carryover |
Some accessories ended up being cheaper than originally budgeted.
|
