Predictors of Ovarian Cancer and Endometrial Cancer for Artificial-Intelligence-Based Screening Tools

Methodology :

This study will involve two different stages

1. The first stage will conduct a cohort study to identify the possible predictors of each type of cancer
2. The second stage will cover the development of point-of-care testing based on an artificial intelligence model to predict cancer occurrence and prospective testing of the new participants using a diagnostic study method. The tools will predict the current histopathology result and possible future histopathology within one year.

Participants and source of data In the study centre, women with or without gynaecology-associated symptoms underwent gynaecological and pathology assessments to rule out ovarian and endometrial cancer in our study centre were involved. Data is stored digitally and extraction will be done accordingly

Variables and outcome measurement

1. Demographic data and health data this information is obtained from the initial assessment of the patients including age, body mass index, chronic diseases, gynaecological and obstetric profile, menstrual pattern, and contraception
2. Clinical and laboratory data this include, a complete blood count, selected cancer-associated biomarker (for example Cancer Antigen 125 (Ca-125)), involvement of lymph node, histopathology of pertinent tissues, and signs of metastases through clinical or radiological data
3. Outcome final histopathology type and classification assessed by at least two pathologists to determine the type of cancer. The guidelines of classification follow the World Health Organization's classification

Development of Artificial-Intelligence-based screening tools

1. The researcher will develop

   • an information-based model where the user will provide a response to each predictor
   • an image-based model where the user will provide a captured image for prediction
   • a mixed-based model where the user can combine captured images and information for each predictor

2. proposed model

   • scoring-based derived from the coefficient of regression
   • decision tree
   • random forest
   • artificial neural network
   • convolutional neural network
3. Selection of model

<!-- -->

1. Screening performance on split data (or using cross-validation technique)

2. evaluation of log-loss or likelihood

   Timeline

   1. For the first stage of the study, there will be a time-varying assessment for each participant, however, at least participants undergo an Assessment of all factors and outcomes at baseline. Repeated evaluation as suggested by the physician will be done within one year after the baseline assessment.
   2. The second study will apply prospective screening. The artificial intelligence-based screening tool will be used concurrently with the gold standard of diagnosis.

   Possible Bias procedural bias particularly in reliability outcome interpretation is handled by involving multiple pathologists. The pathologist and the screener will perform the screening independently to reduce the tendency of prior results provided by the newly-developed screening tools.

   Sample size

   1. The first stage of the research assumes that

a. The prevalence of both cancer among all cancers in women accounted for 5% b. Type I error set at 5% c. absolute error of the prevalence 1% using the one-sample proportion formula, the estimated sample size is 1825 participants.

2. Following the diagnostic study, we state that the new screening tools model will show non-inferiority performance to histopathology as gold-standard, assuming that

a. the expected difference in sensitivity value is 5% assuming that the new screening tools will possess 85% sensitivity and the sensitivity of histopathology is 90% b. cross-over testing will be done, creating an equal allocation of screening intervention c. Type 1 error of the study set at 5% d. Power of the study set at 80% the total sample size for the prospective screening tool will be 1080 participants

Data Quantification and discretization several clinical information will be classified according to the established guideline for example body mass index.

Proposed Statistical Analysis

1. Descriptive statistic and bivariate analysis
2. A cox-regression will be conducted following the baseline-to-event timeline
3. Subgroup analysis will be done, particularly in certain demographic and comorbidity.

as for the second stage, the analysis will identify the

1. sensitivity
2. specificity
3. accuracy
4. precision
5. The number Needed to Treat selected models will be deployed into an application.