Using Machine Learning Techniques to predict malaria prevalence in Rwanda

doi:10.21203/rs.3.rs-5318706/v1

Download PDF

Research Article

Using Machine Learning Techniques to predict malaria prevalence in Rwanda

https://doi.org/10.21203/rs.3.rs-5318706/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Malaria is a terrible communicable disease that leads to the death of people every day throughout the world despite the effort made to eradicate it. This vector-borne disease affects Africa because of limited medical resource, lack of information and other socio-economic factors. In report of WHO 2019, there is an estimation of 228 million of malaria cases globally and in 2023 WHO reported the increase of malaria to 249 million. This increase is associated with many risk factors and those factors mitigated to eradicate malaria. This research aims to predict malaria prevalence in Rwanda. We have managed to reveal factors contribute to malaria outbreak like region, type of place of residence, district, and wealth index, protect against malaria by sleeping under a mosquito net, protect against malaria by sleeping under an ITN, protect against malaria by cutting grass around house, seen or heard any messages about malaria and person slept an ever-treated net. Those factors are associated with the malaria prevalence either positively or negatively.

The logistic regression model is designed to provide future prediction offering valuable insights to mitigate the impact of malaria outbreaks. This approach will enable healthcare providers and authorities to intervene in advance by reducing the severity of malaria and allocate resources where will be mostly needed. This dissertation also provides explanation to the methodology used to develop predictive models where we used data from RDHS from 2019 to 2020 both trained dataset and tested dataset, the sample size was 5041 where we used 16 variables of interest.

Malaria is a deadly disease transmitted by infected mosquitoes. It can be prevented by identifying population at risk and provide proper care. The World Health Organization (WHO) classified it as communicable disease (Cella, 2019). However, Malaria has been declined over the years through various strategies have been initiated to controlling and preventing of malaria such that from 2010 to 2018, malaria incidence has been declined from 71 to 57 per 1000 population globally at risk. In Africa, incidence rate of malaria shifted from 294 in 2010 to 229 in 2018, which constitute 22% of incidence reduction globally. Regarding malaria-related deaths; Malaria is among the primary causes of mortality in the World, where 405,000 deaths caused by malaria in 2018, in 2010 were 585,000 deaths, the under-five year's children are the most vulnerable group because they counted 67% of malaria deaths globally. (Savi M. K., 2022). The most vulnerable groups are pregnant women and children. Despite various malaria control programs in Rwanda, malaria prevalence has increased since 2012 (Karema, 2020) There is an interest in using machine-learning techniques to predict malaria prevalence in Rwanda.

This research considered data from Rwanda Demographic and Health Survey (RDHS) 2019–2020 that carried out by National Institute of Statistics of Rwanda (NISR) from November 09, 2019 to July 20, 2020. It was the nationwide study intended of collecting up-to-date data on demographic and health indicators.

The household Record 1 dataset underwent a thorough cleaning process to deal with data imbalance and missing values by using Python to ensure quality of data and consistency.

We have used Exploratory Data Analysis (EDA) to analyze frequency and distribution of malaria prevalence among population in Rwanda, multivariable binary logistic had used to predict association between independent and dependent variables. we have built machine learning predictive models where we trained and evaluated the household record 1 dataset by using supervised machine learning techniques such as Naïve Bayes, Decision Tree, Random Forest, Logistic regression and K-Nearest Neighbors. The set of codes had written in Jupyter notebook of anaconda. We have also evaluated different models by using confusion matrix and Area under the Receiver Operating Characteristic Curve (AUC_ROC) to evaluate the trade-off between sensitivity and specificity across different classification thresholds. In addition, a synthetic minority oversampling technique (SMOTE) was employed to address the data imbalance problem.

This study established logistic regression model as model with high accuracy to predict malaria prevalence in Rwanda by using Household record1 dataset, with 5041 samples and 1050 features. By using SMOTE, the logistic regression had a highest accuracy of 96% of accuracy and 70% of sensitivity. This study revealed factors associated with malaria prevalence in Rwanda. The threshold for significance was set at p _< 0.05, and the results presented as crude ratios (CORs) or adjusted odds ratios (AORs) and 95% CI. Those features are region, sex of head of household, protect against malaria by sleep under a mosquito net, protect against malaria by sleep under an ITN,protect against malaria by cut grass around house, information about malaria, type of residence, wealth level and marital status (living together) and their odds ration exclude1 in their confidence interval. The most important feature was age of head of household.

Table 1

Distribution of malaria infection and its associated risk factors
Determinants	Levels	Frequency(Percent)		P-Value**
Determinants	Levels	Malaria negative	Malaria positive	P-Value**
Underlying determinants
Region	Kigali city	939(19.49)	106(47.75)	0.000
	South	855(17.74)	23(10.36)
	West	884(18.34)	11(4.95)
	North	1121(23.26)	68(30.63)
	Est	1020(21.17)	14(6.31)
Type of place of residence	Rural	3667(76.09)	197(88.74)	0.000
Type of place of residence	Urban	1152(23.91)	25(11.26)	0.000
Sex of head of household	Female	1885(39.12)	72(32.43)	0.0539
Sex of head of household	Male	2934(60.88)	150(67.57)	0.0539
District	Continuous			0.000
Owns of livestock, herds or farm animals	No	2395(49.7)	115(51.8)	0.5864
Owns of livestock, herds or farm animals	Yes	2424(50.3)	107(48.2)	0.5864
Wealth index combined	Middle	854(17.72)	55(24.77%)	0.000
	Poorer	942(19.55)	58(26.13)
	Poorest	1039(21.56)	64(28.83)
	Richer	868(18.01)	30(13.51)
	Richest	1116(23.16)	15(6.76)
Protect against malaria by: sleep under a mosquito net	No	1969(40.86)	152(68.47)	0.007
Protect against malaria by: sleep under a mosquito net	Yes	2850(59.14)	70(31.53)	0.007
Protect against malaria by: sleep under an ITN	No	3011(62.48)	170(76.58)	0.000
Protect against malaria by: sleep under an ITN	Yes	1808(37.52)	52(23.42)	0.000
Protect against malaria by: cut grass around house	No	1619(33.6)	93(41.89)	0.0132
Protect against malaria by: cut grass around house	Yes	3200(66.4)	129(58.11)	0.0132
Seen or heard any message about malaria	No	1630(33.82)	98(44.14)	0.002
Seen or heard any message about malaria	Yes	3189(66.18)	124(55.86)	0.002
Current marital status	Divorced	201(4.17)	10(4.5)	0.1227
	Living together	1004(20.83)	61(27.48)
	Married	2236(46.4)	96(43.24)
	Never married	410(8.51)	21(9.46)
	Not living together	64(1.33)	4(1.8)
	Widowed	904(18.76)	30(13.51)
Hanged net	No	598(12.41)	4221(87.59)	0.996
Hanged net	Yes	27(12.16)	195(87.84)	0.996
Person slept under an ever-treated net	No	1312(27.23)	80(36.04)	0.0052
Person slept under an ever-treated net	Yes	3507(72.77)	142(63.96)	0.0052
Total		4819 (95.6)	222(4.4)

Identification of important factors of malaria prevalence in Rwanda:

The Table 2 below shows logistic regression results and revealed covariates that are statistically significant. It presents the odds ratios and corresponding confidence intervals and demonstrate influence of some covariates on malaria prevalence in Rwanda. The significance is confirmed when the confidence interval exclude one. This research finding demonstrated a strong association between wealth combined index of household where the richer households are not likely to have malaria but other classes are statistically significant with p-value < 0.05.

Additionally these findings revealed that the possibility of having malaria is significantly associated with region, sex of head of household, protect against malaria by sleeping under a mosquito net, protect against malaria by sleeping under an ITN, protect against malaria by cutting grass around the house, seen or heard any messages about malaria, person slept under an ever treated net, place of residence rural, place of residence urban, current marital status (living together)and wealth index combined.

Table 2

Results of multivariable binary logistic regression
Covariate	Coefficient	P-value	Odds-Ratio	Confidence Interval 2.5%	Confidence Interval 97.5%
Number of household members	0.0954	0.8165	1.1001	0.4912	2.4639
Region	-0.3582	0.0000	0.6989	0.6328	0.7719
Sex of head of household	0.2915	0.0464	1.3385	1.0046	1.7832
Age of head of household	0.0140	0.9683	1.0141	0.5089	2.0206
Owns of livestock herds or farm animals	-0.0841	0.5402	0.9193	0.7023	1.2033
Protect against malaria by Sleep under a mosquito net	0.4056	0.0059	1.5002	1.1238	2.0026
Protect against malaria by Sleep under an ITN	-0.6745	0.0000	0.5094	0.3714	0.6987
Protect against malaria by cut grass around house	-0.3541	0.0111	0.7018	0.5340	0.9223
Seen or heard any messages about malaria	-0.4358	0.0017	0.6467	0.4929	0.8485
Hanged net	0.0229	0.9130	1.0232	0.6780	1.5440
Person slept under an ever treated net	-0.4094	0.0043	0.6640	0.5013	0.8797
Type of place of residence Rural	0.9065	0.0000	2.4755	1.6243	3.7728
Type of place of residence Urban	-0.9065	0.0000	0.4040	0.2651	0.6156
Wealth index combined (Middle)	0.4247	0.0079	1.5291	1.1175	2.0922
Wealth index combined (Poorer)	0.3754	0.0168	1.4556	1.0700	1.9801
Wealth index combined_(Poorest)	0.3878	0.0109	1.4737	1.0935	1.9861
Wealth index combined (Richer)	-0.3408	0.0882	0.7112	0.4807	1.0523
Wealth index combined (Richest)	-1.4253	0.0000	0.2404	0.1418	0.4078
Current marital status(Married)	0.0804	0.8084	1.0837	0.5659	2.0755
Current marital status(Living together)	0.3644	0.0183	1.4397	1.0636	1.9487
Current marital status(widowed)	-0.1277	0.3567	0.8801	0.6709	1.1547
Current marital status(not married)	0.1165	0.6203	1.1235	0.7087	1.7810
Current marital status(Divorced)	0.3099	0.5513	1.3632	0.4920	3.7776
Current marital status(Not living together)	-0.3906	0.0506	0.6767	0.4575	1.0010

Models comparison in machine learning

The Table 3 below revealed the most accurate models, which used to predict malaria prevalence in Rwanda are logistic regression with the highest precision at 96%. The second ones are random forest and KNN with 95%, followed by Decision tree with 91% and the least prediction of the accuracy is Naïve Bayes being the last classifier with precision of accuracy of 77%. To examine the precision we measure how closer the various measurements to each other where we measure precision and find average deviation, to calculate average of the differences in measurements. Data can be precise without being accurate but ideally measure the precision and accuracy. As the precision measure the correct predicted values belongs to specific factor class compared to all values, where Naïve Bayes classifier exhibit the model with 94% followed by Random forest and decision tree with 93%. The KNN and Logistic regression have the precision of 92% of the model.

The F1-score is the inverse of combination of the recall and precision; the random forest, KNN and logistic regression prove predicting power with 94% based on our data as F1-score is high; It demonstrates a fair balance between precision and recall of the classifiers. It means that the model performed well on both aspects. Regarding trained and tested datasets, logistic regression considered as the best predicting classifier with high predicting accuracy, recall and f1-score and Gaussian (Naïve Bayes) is the least predicting classifier with lower prediction in predicting accuracy, recall and fi-score.

Table 3

Results of multivariable binary logistic regression
	Model	Precision score	Recall score	F1_score	Misclassification	Accuracy	Cohen Kappa
0	Logistic Regression	0.9248	0.9616	0.9429	0.0383	0.9616	0.0
1	Naïve Bayes	0.9401	0.7714	0.8403	0.2285	0.7714	0.0789
2	Random Forest	0.9335	0.9577	0.9431	0.0422	0.9577	0.0480
3	Decision tree	0.9368	0.9114	0.9233	0.088	0.9114	0.1300
4	KNN	0.9246	0.9564	0.9402	0.0435	0.9564	-0.0093

There has been a lot of interest about factors driving the persistence of malaria in Rwanda despite the various intervention and control measures that have been put in place. The features associated with malaria prevalence were assessed. This study revealed that region, sex of head of household, protect against malaria by sleep under a mosquito net, protect against malaria by sleep under an ITN, protect against malaria by cut grass around house, information about malaria, type of residence, wealth level and marital status (living together) are associated with malaria.

Different studies had identified risk factors such as socio-demographic, climatic and environmental conditions and numerous important methods to understand its prevalence and associated risk factors, Several studies have been performed to assess the prevalence of malaria among children under the age of five, and it is widely recognized by socio-demographic factors significantly influence the occurrence of malaria (Adamu, 2021).

The country conducted an in-depth national data analysis to ascertain the potential causes of the increase in cases and design appropriate malaria control interventions. Here are some reasons identified that are Climate change, mosquito behavior change, lower interventions in high impact zone, unreliable vector control activities, and increases malaria cases and reporting rates from health facilities and the community into the system (Cella, 2019).

(Bridget, 2021) Revealed that the effect of climate variability is the one leading factors influence malaria prevalence across geographical locations where she conducted research to predict malaria incidence in sub-Sahara Africa as we find that in Rwanda also the prevalence relate to the region, place of residence and district.

As presented statistically significant variables (p-value < 0.05) include region, gender of household head, malaria prevention methods (e.g. mosquito nets, ITNs, cutting grass around the house), exposure to malaria-related messages, type of residence, wealth index, and current marital status (living together), with odds ratios excluding one in the confidence interval.

In this work, we evaluated the performance of several algorithms, namely the degree and accuracy of malaria prediction algorithms capable of detecting malaria outbreaks in Rwanda. Performance assessed by using the dataset's accuracy, Area under Curve (AUC), precision, sensitivity or recall, specificity, F1-score, True Positive Rate (TPR), False Positive Rate, Average Macro, and Average Weighted components. Random forest performed far better than other classifiers.

The best three performing classifiers in our study are random forest, decision tree and K-NN. The limitation of this research is that we did not include the climate data or other environment factors and we have considered socio-economic characteristics and geographical features as are the ones provided in RDHS to assess their impact on malaria prevalence.

Additional factors such as proximity to marshlands, irrigation schemes, and cross-border movement of people influence the transmission, particularly in the country's southern and eastern regions (WHO, 2022).

The factors that are linked to human actions can change the environment that read to increase of the transmission compare to natural factors like natural disasters, climatic variations, or any other suddenly modification of environment and cause a large-scale population movements that may resurgence of malaria epidemics as control services becomes deficient and no presence of control measures.

This study was based on RDHS data from 2019 to 2020, the results revealed that the age of head of household, number of household members, protect against malaria by using different measures like cutting grass around house, sleep in under an ever-treated net, region and access of information etc. were the very importance features in machine learning. This study revealed the most significant determinant of malaria occurrence in different parts of the country that are region, use of mosquito net and access to information; and the machine learning algorithms with highest precision in predicting malaria prevalence is logistic regression with 96% of accuracy. The more accurate prediction is essential in determining where the outbreak will occur. This lead to the intervention of health providers and government in advance to take necessary measures to ensure the safety of the population in a given area.

Author Contribution

All authors reviewed the manuscript

Adamu Y. &. (2021). Malaria prediction model using advanced ensemble of machine learning techniques. J Med Pharm Allied Sci.
Alonso PL, Brown G, Arevalo-Herrera M, Binka F, Chitnis C, Collins F, Tanner M. (2019). A research agenda to underpin malaria eradication. PLoS Med, 8(1), e1000406.
Bhatt S, Weiss DJ, Cameron E, Bisanzio D, Mappin B, Dalrymple U, Gething PW. The effect of malaria control on Plasmodium falciparum in Africa between 2000 and 2015. Nature. 2015;526(7572):207–11.
Bridget ON. (2021). Machine Learning Techniques for Malaria Incidence and Tuberculosis Prediction.
Brownlee J. (2020). Machine Learning Mastery.
Caminade C, Kovats S, Rocklov J, Tompkins AM, Morse. A. P., Colón-G.
Cella W. (2019). Do climate chnages alter the distribution and transmission of malaria?.
Karema C. (2020). History of malaria control in Rwanda: implications for future elimination in Rwanda.
Savi. &. M. (2022). An overview of Malaria Transmission Mechanisms, Control, and Modeling.
WHO. (2022). World Malaria Report. WHO.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Using Machine Learning Techniques to predict malaria prevalence in Rwanda

Status:

Version 1

Abstract

Background

Methods

Results

Identification of important factors of malaria prevalence in Rwanda:

Models comparison in machine learning

Discussion

Conclusion

Declarations

Author Contribution

References

Additional Declarations

Status:

Version 1