The data for this study was gathered from 7 published papers, shown in Table 1, that perform random clinical trial to investigate the intermittent fasting effects on T2DM parameters.

The collected data included 838 individuals and 13 different types of interventions. The following interventions: CER, High Carbohydrate , High Monounsaturated, DR70 and CCR included restricted calorie diets or specific food compound' diets but did not include fasting at all , since they were used for control in the published studies of the random clinical trials. Those interventions were included in the data of this study since they provide interesting information to analyze. For example, in cases where none of the fasting methods improve the T2DM risk parameters of an individual while a calorie restriction did improve. The fasting interventions based on weekly days , which included in this study were Intermittent energy restriction - 2 days a week trail and eating free in all the other 5 days, Fasting Every Second Day; eating only four days a week, Intermittent Energy and Carbohydrate Restriction; eating restricted calories only two days a week, Intermittent Energy and Carbohydrate Restriction plus free Protein and Fat; eating restricted calories only two days a week, Fasting three non-consecutive days per week and Fasting three non-consecutive days per week and on eating days have 70% energy. The Daily Morning Fasting is a fasting intervention based on day's hours. In total 838 Individuals records were collected from 13 different intermittent fasting interventions. The distributions of the 13 different interventions are shown in Figure 1.

This study focuses on classification, determining if specific interventions improve HOAM-IR for pre-diabetes and diabetes individuals. Four classifiers were chosen: decision tree (J48), logistic model tree, random forest, and logistic regression. Decision trees are simple and interpretable, while J48 avoids overfitting. Logistic model trees handle complex relationships, and random forests excel with high-dimensional data. Logistic regression is straightforward. These classifiers were selected due to their advantages and compatibility with the dataset. Decision trees handle missing values, random forests are accurate in high dimensions, J48 avoids overfitting, logistic model trees manage complex relationships, and logistic regression is easy to implement.

The test approach which is used is 10-fold which is building the model using 9 over 10 of the data and testing with 1 over 10. This is done 10 times.3. Results

This section may be divided by subheadings. It should provide a concise and precise description of the experimental results, their interpretation, as well as the experimental conclusions that can be drawn.

The initial step in selecting the optimal intervention for a patient with reported prediabetes is to address the question: does a specific intervention improve HOMA-IR? Four machine learning classifiers were chosen to answer this question: J48 decision tree, Logistic Model Tree (LMT), Random Forest, and Logistic. Detailed information about each classifier can be found in the Method section. The dataset described in the Methods section was trained and tested, utilizing the 10-fold test to assess the predictive ability of each classifier for HOMA-IR improvement. Improvement in HOMA-IR was defined as a decrease in HOMA-IR after the intervention compared to before the intervention. The results for each classifier can be found in Table 2 under the "Discrete difference" row and "HOMA-IR (with control)" column. The results show modest significance, with AUC values ranging from 0.6 to 0.7 (J48: 0.68, LMT: 0.6, Random Forest: 0.68, and Logistic: 0.7), and accuracy ranging from 70% to 72% (J48: 70%, LMT: 72%, Random Forest: 70%, and Logistic: 71%). To reduce data noise, 61 records of control interventions (CER, DR70, and CCR) were removed from the dataset. These control interventions did not involve fasting. After removing these records (see Methods section for details), slight improvements in prediction were observed for all four classifiers. The results can be found in Table 2 under the "Discrete difference" row and "HOMA-IR (no control)" column. The significance remained moderate, with AUC values ranging from 0.65 to 0.71 and accuracy ranging from 68% to 74%. Interestingly, when the difference between post-intervention and baseline HOMA-IR increased, the results became more significant in terms of AUC and accuracy. In this case, improvement in HOMA-IR was defined as a decrease in HOMA-IR after the intervention by more than 15% compared to the normal HOMA-IR (calculated using fasting glucose of 5.56 mmol/liter and fasting insulin of 80 pmol/liter; calculation explained in the Methods section). Using a cutoff of greater than 15% decrease, the results are presented in Supplementary Tables S3A, S4A, and S5A for different fasting glucose and HOMA-IR differences (before and after interventions) with cutoffs of 15%, 10%, and 20% respectively. The supplementary tables demonstrate that the 15% cutoff performs the best. The results for each classifier, using the cutoff of 15%, can be found in Table 1 under the "Discrete difference above 15%" row and "HOMA-IR (no control)" column. AUC values ranged from 0.73 to 0.89, and accuracy ranged from 74% to 82%. The Logistic Model Tree classifier exhibited the most favorable results, with an AUC of 0.89 and an accuracy of 82%. These findings suggest that the proposed method can serve as a robust foundation for a recommendation system. Table 2 shows that using the Logistic Model Tree classifier, HOMA-IR improvement can be predicted with an accuracy of 82% and an AUC of 0.89. The other three classifiers also provided similar predictive results.

Establishing a solid foundation for a recommendation system requires assessing the predictability of HOMA-IR improvement and the accuracy of the predictions. To determine this, we can leverage our dataset without the intervention feature. Additionally, it would be intriguing to investigate the decision rules for HOMA-IR improvement by training and testing our data without the intervention column. The results of this test, using the four classifiers, are presented in Table 2 under the "Discrete difference above 15%. No Interventions" row and "HOMA-IR (no control)" column. The AUC values range from 0.71 to 0.88, with accuracy ranging from 72% to 83%. While these results are slightly less significant than when utilizing the intervention column (AUC between 0.73 and 0.89, accuracy between 74% and 82%), it is reasonable since omitting the intervention column leads to some information loss. Nonetheless, these results are meaningful and support the recommendation system. For instance, the logistic model tree classifier demonstrates an accuracy of 83% and an AUC of 0.88, while the Random Forest yields an accuracy of 79% and an AUC of 0.86. Interestingly, Figure 2 highlights the significance of fasting glucose and fasting insulin as influential factors in predicting HOMA-IR improvement using the J48 classifier. Furthermore, the figure reveals the substantial role of BMI in determining the appropriateness of IF for reducing HOMA-IR in individuals below 60 years of age. In contrast, for individuals aged 60 and above, gender, specifically women, appears to have a higher likelihood of benefiting from an IF approach to reduce HOMA-IR.

In real life, patients typically undergo only fasting glucose blood tests as part of their routine health check-ups, while fasting insulin tests are not commonly included in the standard blood tests conducted by Health Maintenance Organizations (HMOs). Therefore, a recommendation system based solely on fasting glucose levels would have lower performance compared to a system based on HOMA-IR. However, it is still intriguing to evaluate the performance of our data when using only fasting glucose. The results of the glucose test can be found in Table 2 under the "Discrete difference above 15%. No Interventions" row and "Fasting Glucose (control)" column. Interestingly, the accuracy of the prediction ranges from 93% to 95%, compared to 74% to 77% with the same test for HOMA-IR. Additionally, the AUC values show improvement, ranging from 0.64 to 0.91, compared to 0.75 to 0.83 for the HOMA-IR test. The results of the glucose test without control or interventions, as shown in Table 2, are similar to the HOMA-IR tests but with better accuracy and AUC values. The higher accuracy when using only fasting glucose can be attributed to the smaller number of successful cases in improving fasting glucose (24 out of 474), which is lower compared to the number of cases with improved HOMA-IR (150 out of 471). However, the improvement in AUC is not affected by these numbers, indicating that the data contains relevant information.

The random forest classifier performs best for the dataset using only fasting glucose, where the class is considered True if the difference between basal and post-intervention is higher than 15% of normal fasting glucose. In Table 2, under the "Discrete difference above 15%" row and "Fasting Glucose (no control)" column, the random forest classifier achieves an AUC of 0.93 and an accuracy of 96%. The J48 classifier, although slightly less accurate than the random forest, has the advantage of providing a visible decision-making process. Figure 3 illustrates the decision tree based on the J48 classifier using only fasting glucose without the intervention columns but with control. Six decision pathways are depicted, with four resulting in a False decision and two leading to a True decision. Lower basal fasting glucose tends to lead to a False decision, while being a female with lower weight favors a True decision. Conversely, being a female with higher weight leans towards a False decision. Additionally, older males tend to result in a False decision, whereas younger males tend to result in a True decision. These results are logical and allow us to conclude that weight is the crucial feature for females, while age is the determining factor for males.

The previous sections extensively discussed the ability and success in predicting whether an intervention would improve T2DM risk parameters for a patient with reported pre-diabetes. The next crucial question to address in building a recommendation system is determining the most effective intermittent fasting method for an individual. This question can be answered using continuous difference analysis. In this approach, the target column represents the difference between the basal fasting glucose level and the fasting glucose level. In previous analyses, the target column was labeled as TRUE when the basal fasting glucose level was higher than the fasting glucose level; otherwise, it was labeled as FALSE. However, with the continuous column as the target, we can select the method with the highest difference as the best match. By using a continuous column as the target, the solution shifts from binary classification (True or False) to predicting the value of the difference. The success of this prediction can be measured using the correlation coefficient between the real and predicted values. For this prediction, the random forest algorithm was employed. The results can be found in Table 2 under the "Continuous difference" row and "Fasting Glucose (control)" column, with the random forest algorithm achieving a correlation coefficient of 0.51. The significance of these results will be discussed in the next section, where the correlation coefficients of these algorithms will be compared with a random dataset.

The training set for our recommendation system utilized the original Fasting Glucose dataset with interventions and without control, with the class column replaced by the difference between the basal and post-intervention fasting glucose levels. The test set consisted of 13 records with the same features: age, gender, weight, BMI, and basal fasting glucose level. However, the intervention feature varied across the 13 records since our study included 13 different interventions (without control). As an example, let's consider a pre-diabetic woman who is 54 years old, weighs 73 kg, has a BMI of 26.7, and a basal fasting glucose level of 6.1 mmol/L (110 mg/dL). Inputting this data into our recommendation system yields the output shown in Table 2 under the "Continuous difference Random Forest" column. The random forest algorithm recommends the 'IECR+PF' intervention as having the largest difference, making it the best intervention for this case. Two different cases for men are shown in Tables S1 and S2, where S1 recommends the 'diet high mono' intervention according to the random forest algorithm, while S2 suggests the 'CCR' intervention.

An interesting and essential question in prediction queries is whether the same results can be obtained randomly. To address this, we utilized the same dataset but with a randomized target column while maintaining the proportion between TRUE and FALSE labels identical to the original target column in each test. The results of all random tests can be found in Tables S3-S6 in the supplementary materials. Notably, all random tests consistently exhibited significantly lower success rates compared to the real tests. For instance, Table S3A presents the results of the dataset with a target column indicating a 15% difference, while Table S3B displays the random results for the same dataset. In the real test, the AUC for fasting glucose without control and with intervention ranged from 0.82 to 0.93, whereas in the random test, the AUC ranged from 0.47 to 0.57. Accuracy in the random test is not informative since the TRUE and FALSE labels are not equally distributed in the target set. Therefore, if the TRUE labels account for 20% of the data, we could achieve 80% accuracy by predicting FALSE for all cases. Table S6 demonstrates the random test for continuous difference, where the target column was replaced by random continuous values within the range of the highest and lowest differences found in the original column of the dataset (-2 to 2 mg/dL). As seen in Table S6, the correlation coefficients for all random tests are significantly lower than the results of the original test. These outcomes emphasize the fact that the successful predictions described in the previous sections cannot be achieved randomly.

The following supporting information can be downloaded at the website of this paper posted on Preprints.org, Table S1-S2: examples of recommendation system's results. Table S3: Different cutoff values for difference 15%. Table S4: Different cutoff values for difference 10%. Table S5: Different cutoff values for difference 20%. Table S6: Random test for continuous difference.