An Artificial Intelligence Approach for Data Modelling Patients Inheritance of Sickle Cell Disease (SCD) in the Eastern Regions of Saudi Arabia

Sickle cell disease (SCD) is a rare heredity blood disorder disease which was recognized in 1963 specifically in the eastern region of Saudi Arabia. Most SCD patients live in the eastern province where, of every 10,000 people, 145 are SCD patients. The south-eastern province has a smaller number of patients: of every 10,000, 24 are SCD patients. Sickle cell trait (SCT) carriers have been observed in Saudi Arabia: approximately 2% to 27% of the Saudi population [1] across different regions. Among children aged 5 to 14 years, SCD has the worst results with highest number of patients, followed by children and infants under 5 years of age [2]. Consanguineous marriage (marriage between closely related relatives) among 57.7% of the population was a key reason for these statistics, and this percentage could easily be increased to over 80% in countries such as Saudi Arabia [3]. Furthermore, when two parent have SCT, they have a 25% chance of transmitting their child with the disease; the percentage is even higher (50%) when one parent has SCT and the other has SCD. Saudi Arabia has a significant challenge controlling the disease because the community is not sufficiently educated on the consequences of consanguineous marriages. Although the government implemented programs to educate people and decrease the risks of SCD patients from marriage contracts using medical tests in a premarital screening program, the results indicated couples continued to practice with marriage procedures and local traditions.

The existing techniques in the literature have been investigated in these regards and different approaches have been put into trial especially the statistical analysis [1]. Each approach has its own pros and cons. The current study is novel and unique in the following way.

(1) Nature of dataset: the dataset is novel and no one else has investigated it in the past. It is based on realistic data from KFU Hospital spanned over 12 years clinical records. Moreover, the dataset is versatile and contains several features that none of the existing studies focused.

(2) Earlier studies investigated limited number of algorithms, usually one or two also lacking AI based approaches. While the undergoing study investigated several AI based classifiers over the unique dataset to appropriately analyze the feature set.

(3) The study is focused on Saudi Arabia’s eastern region (Dammam).

This research aims to employ artificial intelligence (AI) modeling techniques through data mining to understand the prevalence of SCD, determine the inheritance information among SCD patients, and study the hospital admissions and discharge causes for SCD. Experiments were conducted on five most frequently used classifiers on SCD, namely naïve Bayes, J48, SVM, neural network (NN), and PART, on SCD patient’s clinical data and relevant knowledge was extracted. These five models were carefully chosen based on review of the literature work [4] conducted as part of ongoing research.

This section comprehensively details the retrospective study conducted as a first step on each of the target questions for SCD patient data through descriptive statistics [6]. The study aimed to understand the determinant factors associated with SCD patients who are getting admitted and being treated at different departments of the hospital.

3.1 Patient epidemiology

While SCD is a rare blood disorder worldwide, incidence of this disease in Saudi Arabia continues to increase mainly because consanguineous marriages are still practiced, resulting in children inheriting the disease and ultimately suffering [7]. In the eastern region of the country, disease statistics are particularly high despite government efforts to reduce disease prevalence over the past decade [8]. Children could inherent SCD from their mother or father and have a 50% risk of inheriting the disorder if one of their parents has a sickle cell gene. As there are little or no signs, sickle cell trait (SCT) may not be discovered early; hence, patients even with 50% transfer SCD to their child if both parents have a sickle cell trait [9]. In this study, we discovered 100 unique patients (children) who inherited illness from their parents (mother).

Figure 1 indicates the ages of those patients who were admitted and got treated in the hospital along with the historical SCD records of their mothers. As can be seen, patients (children) with inheritance details available were between the ages of 1 and 25 years; majority of these children were between 10 and 20 years of age. Newborns had the least patient inheritance track record information.

1.png

Figure 1. SCD inheritance by infants and children

3.2 Hospital visits

The analysis on hospital visits was performed on SCD patients who visited between Jan 2008 and Apr 2020. As can be seen from Figure 2, the number of hospital visits clearly sparked on SCD cases over the past decade. The most patients under the age of 25 with SCD were singles, while married patients were mainly over the ages of 20. For patients between 20 and 40 years of age, more appointments were frequently made.

Table 2. Population visits statistics and marital status

Table 2 provides frequency statistics of hospital visits made by SCD patients, along with their ages and marital status (married, single, not specified). We discretized the number of hospital visits within the last 12 years into five categories based on visits (low: visits below 50; moderate: visits between 50 and 100; over moderate: visits between 100 and 500; high: visits between 500 and 1000; and very high: visits over 1000). The low visits group had 868 patients (of 1428) with different marital status.

The number of married low visit patients were 330 (38.02%) and 15 (1.73%) for single patients. In exceptional conditions, three married patients and five single patients visited the hospital more than 1000 times. Overall, among all genders, mean ages of married patients who visited the hospital was 41 and, patients who are singles was 23 years.

2.png

Figure 2. SCD hospital visits by different marital status

3.3 Gender demographics

Table 3. Gender and inheritance statistics of SCD patients

Table 3 displays the residential cities of all the patients, along with gender and inheritance statistics, who live throughout Saudi Arabia. The results indicate that there were 58% female patients with SCD (759) than 52.17% male patients (669) who visited the hospital living mostly in and around Al Khobar city. Furthermore, 21.61% were females from Al Qatif town while 15.70% were SCD males. In addition, cities in the eastern regions of the country had more SCD patients. For instance, 10.67% female patients and 9.12% male patients were from Dammam city. Al Ahsa also has lower female patients (7.11%), while Dhahran and Jubail had lower male patients. In addition, 16% of patients came from other regions, such as the city of Riyadh, the capital city.

Table 3 presents the inheritance statistics. The number of patients who inherited the disease were 100 (7%) among 1428 patients. Among these patients, 75% lived in Al Khobar city, and other SCD patients lived in other eastern region cities: Al Qatif (8%), Dammam (8%), Dhahran (6%), and Al Ahsa (2%). We determined the relationship between genders and patient ages as shown in Figure 3 and found that female patients were greater than male patients over 30 years old. Almost 50% of both female and male patients were between the ages of 30 and 40. Less than 5% of patients were over 60 years old.

3.png

Figure 3. SCD patients gender (male/female) statistics

4.png

Figure 4. Top 5 services (visits) by each SCD type

3.4 Treatment types

SCD patients received various clinical services from different hospital departments. The services received by patients based on various forms of SCD are shown in Figure 4. As can be seen, with a combined (SCD, SCT, and SCC), the most visited department in the hospital found to be Internal Medicine for Adults, followed by combined visits at the Department of Pediatrics for children [10]. For SCC patients, the top 5 services received in the descending order are the Internal medication (110,570), followed by pediatrics (13,986), cardiology (5530), gynecology (5377), and hematology (5317). For SCD patients, pediatrics (17,933) was the highest, followed by internal medication (4401), pediatrics (3861), gynecology (1909), and general surgery (1632). For SCT patients, though minimal visits, internal medication (513) was still the highest, then gynecology (265), pediatrics (157), gastroenterology (59), and orthopedic surgery (53).

5.1 Neural networks

Classification was performed with default values for all NN parameters except the “hidden layer” [17]. Each time the experiment was done but with different types and numbers of hidden layer to find the best combination classified number of instances. Figure 5 shows a generic NN architecture for binary classification with five layers, including input, hidden and output layers.

5.png

Figure 5. Generic NN model for binary classification

6.png

Figure 6. Accuracies for different hidden layer values

The values for NN parameter set in the experiment are letters: “a” (default) is the dataset’s classes and attributes divided by 2, “o” is the number of classes in the dataset, “t” is the dataset’s classes plus the attributes, and “i” is the number of attributes in the dataset. Figure 6 represents different hidden layer types with different accuracies. The hidden layer value “a” (default value) had highest classification accuracy, so we choose “a” as optimal value. Experiments were repeated with different seed values 1, 100, 200 and 300 and hidden layer was changed to “a”. Our results (as shown in Figure 7) revealed best accuracy of 92.99% is achieved when the seed is set to 1.

7.png

Figure 7. Accuracies for different seed values

Finally, the last parameter manipulated was learning rate, which had different results when all parameters remained at the default value, except for the hidden layer “o” and seed to 1. Figures 8 and 9 display the accuracies of different learning rate values: 0.3 had the highest accuracy result at 92.99%. Table 4 summarizes experimental results on the optimal values obtained for the NN classifier on the patient’s data model.

8.png

Figure 8. Different learning rates with accuracy results

9.png

Figure 9. Accuracies of different kernel types

Table 4. Optimum parameters of NN

5.2 Naïve bayes

The classification accuracy of the Naïve Bayes [18] classifier is primarily affected by two factors. The first parameter is the numeric value’s distribution, which is set to Gaussian by default but may be changed to kernel by changing the “kernel estimator” from False to True. The “supervised discretization” option is the second parameter, which allows the user to convert numeric values to nominal values. The two factors had no influence on the patient’s clinical data model for Naive Bayes. As a result, the default parameters were used as the optimal values.

5.3 Support vector machine

The parameter optimization strategy was employed to the patient clinical data model using Support Vector Machine (SVM) [19]. It is among the promising classifiers in healthcare data analytics. First, the experiment used all default values for the SVM parameters, except for the kernel parameter, each time the experiment was run with a different kernel value. Figure 8 provides the accuracies for different kernel types and highlights that the radial basis function and polynomial had the highest accuracies at 92.99%. Since the two kernel types had the same accuracy result, we chose the default, polynomial. The second parameter, cost, was changed many times (1, 10, 50, and 100), but all other parameters remained with default. Figure 10 displays the variation of the accuracy results for different cost values. The highest accuracy was 1 at 92.99%.

10.png

Figure 10. Accuracies of different cost values

Table 5 presents the optimum values of the proposed SVM algorithm for the patient clinical data model.

Table 5. Optimum parameters of SVM

5.4 J48

The J48 classifier was applied as part of the optimization strategy to find the best accuracy result. It is quite effective probabilistic and lightweight classifier for binary problems. Figure 11 provides different values for the confidence factor and their accuracies. As can be seen, the highest accuracy result of confidence factor values among 0.1, 0.25, and 0.4 scales were achieved at 92.99%, so the optimal value was set to 0.25, the default value for the confidence factor.

11.png

Figure 11. Accuracies of different cost values

Table 6. Optimum parameters of J48

The manipulation of other parameters did not affect the accuracy of the J48 classifier, which remained at 92.99%; all other parameters were optimal with their default values. Table 6 presents the summarized optimum values detected for J48 parameters on the patient clinical data model.

5.5 PART

The PART classifier was employed as part of optimization strategy to find the best accuracy result. Figure 12 depicts different values of minimum number of objects and their accuracy results that were experimented. The accuracy result indicates that 8 is the most optimal value achieved for minimum number of objects at 92.92%. The second parameter was seed and its values were changed every time the experiment was run, but it always provided the same result (92.92%), which is the default optimal value. Table 7 provides the optimal values of the PART classifier for the patient family inheritance model.

12.png

Figure 12. Accuracies of PART classifier

Table 7. Optimum parameters of PART

Sickle cell disease, a blood disorder, is widespread among people primarily living in the eastern regions of Saudi Arabia due to the common practice of consanguineous marriages. Since there is no permanent cure, researchers and doctors have developed temporary methods to make life easier for SCD patients and their families. In this research, we first conducted statistical retrospective analysis and predicted the problems and various factors associated with SCD. We visualized and discussed in detail the patients’ disease inheritance reasons currently occurring in regions where the disease incidence is getting higher than the rest of the nation.

King Fahad University Hospital is a major public hospital located in the eastern province region. Through serving SCD patients from different cities, doctors at the hospital have noticed the rapid growth of SCD cases over the past decade. This research involved employing prediction models on twelve years (2008-2020) of SCD clinical data, obtained through official ethical approval process. This involved applying the five unique classification models most frequently employed for SCD research (naive Bayes, NN, SVM, J48, and PART classifiers) which were identified through the review of literature conducted in our previous work. The experiments involved analyzing, predicting, and identifying the optimize parameters of classifier algorithms. We measured and found naïve Bayes outperformed in the experiments (92.22%) which was further selected to extract the needed SCD patients’ inheritance knowledge (primarily from their mothers).

In conclusion, this research provided critical knowledge on SCD inheritance related causes, and management strategies in the eastern region of the country where SCD cases are rapidly growing especially in the last decade. Knowing data relationship between the inheritance of the disease and other factors would allow the government to educate and manage the disease by increasing disease awareness in the community, especially among mothers. The future direction of this research could consider how SCD patients who are infected with the currently ongoing pandemic COVID-19 is a major area of interest. The future research direction could also include how this rare heredity disease is infecting behavior of SCD patients living under different living conditions.