Development and Evaluation of a Predictive Diagnostic System for Dengue Fever using Machine Learning Techniques

doi:10.21203/rs.3.rs-2473833/v1

Download PDF

Research Article

Development and Evaluation of a Predictive Diagnostic System for Dengue Fever using Machine Learning Techniques

https://doi.org/10.21203/rs.3.rs-2473833/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this older preprint version

Read the latest preprint version →

As the world continues to evolve scientifically and climatically, the medical and healthcare sectors face new challenges in the form of newly identified diseases, insects, chemicals, and environmental effects. These new challenges are often more complex and dangerous than existing conditions, and require prompt diagnosis and appropriate treatment to prevent loss of life. This paper presents a study on the development of a presumptive diagnosis system (PDS) for early dengue fever detection. The proposed system aims to assist medical professionals in determining the disease at an early stage and help unprofessional or new staff to learn and experiment better. The PDS utilizes machine learning algorithms such as Decision Tree (DT), Random Forest (RF), and Naive Bayes (NB), along with their combination, to classify dengue fever and its types. The dataset used in this study consisted of 400 records of 15 attributes, and the data was preprocessed to reduce noise, incompleteness and inconsistencies. The classification results showed that the PDS achieved an average accuracy of 95.6%, with the highest individual accuracy achieved by RF, DT and NB being 98%, 97% and 87% respectively. Our results demonstrate that PDS can facilitate the diagnostic process in resource-limited settings by providing an alternative to hardware-intensive solutions. Furthermore, the system can aid healthcare professionals in making early and accurate diagnoses, and can also serve as a learning tool for new staff.

Dengue fever

Healthcare

Early diagnosis

Predictive modeling

Disease is a dysfunction or an abnormal condition in living organisms that impairs their normal functioning. There are a variety of diseases that can affect living organisms, including viruses, bacteria, parasites, and more. Malaria and dengue fever are specific examples of diseases that are caused by the transmission of viruses and parasites through the bite of infected mosquitoes. These diseases are known as mosquito-borne infectious diseases. Both malaria and dengue fever have similar symptoms[1], [2], that appear after 3–14 days of an infected bite[3]. Symptoms usually include fever, headache, muscle pain, fatigue and rashes. Malaria is caused by the Plasmodium parasite, which is transmitted to humans through the bite of an infected Anopheles mosquito. It affects more than 200 million people worldwide and causes more than 400,000 deaths each year[4]. On the other hand, Dengue fever is caused by the Dengue virus, which is transmitted by the Anopheles Mosquitoes. It is a leading cause of illness and death in many tropical and subtropical regions of the world.

One of the main challenges with malaria and dengue fever is that their symptoms often only appear a few days after the initial bite, and if not treated properly, can lead to severe illness or death. According to the World Health Organization (WHO), there are over 900 million cases of malaria[2] and dengue fever reported each year, with many countries affected. It reported that 390 million people get infected by dengue yearly[1], and many could not survive and 350 to 500 million cases of malaria reported each year. According to a WHO, 3.9 billion people are at risk in 129 countries globally of contracting dengue virus, with 70% of cases concentrated in Asia[1], [3], [5].

Malaria and dengue fever are serious health concerns that affect millions of people worldwide. These diseases can cause severe illness or death if not treated properly. To address this issue, the use of Presumptive Diagnosis System (PDS) is crucial to detect the disease in early stages and assist less experienced staff and healthcare workers in enhancing their understanding of the disease and decision making, would allowing them to take the necessary measures to prevent further spread of the disease and ultimately save lives.

The literature work discussed various applications of classification algorithms and machine learning techniques in various fields such as medical diagnosis[6]–[8], substance abuse[9], and diabetes[6], [8], [10], [11] prediction. In the study[12], authors used anthropometric data and blood analysis parameters to classify patients with breast cancer using artificial neural networks and Naive Bayes classifiers, with high potential for accurate diagnosis. The psychosocial factors[9] such as life events, aggression behavior, and interpersonal trust have a greater impact on substance craving in male substance users. The authors in[10] introduced a new labeled diabetes dataset from Bangladesh and proposed an automated classification pipeline using a weighted ensemble of machine learning classifiers, which produced the best results for early diabetes prediction. In another literature, proposed a method for measuring empathy[13] in text-based conversations using a pattern-based classification algorithm, and found that an explicative pattern-based approach (PBC4cip) is the best approach for detecting empathy in texts. In another study[14], investigated the use of natural herbal based antimicrobial agents in textile materials and found that methanol extracts showed better antimicrobial activity. While [7], [11], studied the utilization of soft computing techniques in diagnosing tropical febrile diseases, and found that ensemble techniques were used more frequently than single machine learning techniques, with Dengue fever being the most studied disease. The other works[7], [10]–[13] study aimed to address the diabetes problem effectively by utilizing artificial intelligence-based detection techniques, such as density-based spatial clustering[15] of applications with noise and ordering points to identify the cluster structure, machine vision systems, and iridocyclitis for detection of the disease through iris patterns.

The literature discussed the use of Information and Communication Technologies (ICT) in healthcare, specifically the use of machine learning and data mining techniques to improve the diagnosis and treatment of various diseases. It highlights the ability of ICT tools to analyze large amounts of medical data and identify patterns that can assist in early detection of diseases, and in making better-informed treatment decisions. Additionally, the integration of ICT in healthcare opens new opportunities for remote monitoring, telemedicine, telehealth and artificial intelligence in medical decision making, which can improve accessibility and reduce the need for patients to travel long distances to receive care.

Machine Learning Algorithms

Decision Tree is a method used for classification, data manipulation, interpretation, and prediction. It creates a tree-like representation of data to predict outcomes based on decision variables. It is easy to understand, handles complex datasets, and works with discrete and categorical data.

Random Forest combines multiple decision trees to improve results and classify new instances using internal voting. It uses bootstrap sampling and Out of Bag (OOB) error. It is commonly used in various fields including Medical diagnosis.

Naive Bayes is a simple technique for classification in machine learning, which uses probability to predict class based on Bayes theorem and total probability theorem. It is easy to construct and performs well with large datasets. Despite its simplicity, it often performs well in practice.

DATASET

The dataset used in this work is collected from the authors [16], contain the data collected from various hospitals in Lahore, Pakistan during a dengue outbreak. The data includes patient's history, examination results, and laboratory test reports. The dataset used in this study is collected from various hospitals in Lahore, Pakistan, and includes information on symptoms and lab test results of patients diagnosed with dengue fever. The data includes 5 numerical attributes, 14 categorical attributes, and 1 target attribute. The dataset includes 376 instances, with 101 instances of dengue fever, 73 instances of dengue hemorrhagic fever, 97 instances of dengue hemorrhagic fever type 1, and 99 instances of dengue hemorrhagic fever type 3. To improve the quality of the data mining process, the data was preprocessed by cleaning, integrating, and transforming the data. This resulted in the removal of irrelevant or redundant attributes and the extraction of hidden or missing information. Finally, the Table 1 the subset of 15 relevant attributes out of the original 20 attributes.

Table 1

Dengue Dataset type of Attributes– categorical, numerical as well as class or target
S#	Attribute Name	Meaning	Data Type	Data Class	Values
1.	Fever	High grade temperature lasts for 2–7 days	Text	C	Y/N
2.	Headache	Severe headache	Text	C	Y/N
3.	Retro orbital pain	Red eyes or pain behind eye- ball	Text	C	Y/N
4.	Muscle pain	Deep Muscular pain during illness/ Myalgia	Text	C	Y/N
5.	Joint Pain	Pain in all joints /Arthralgia	Text	C	Y/N
6.	Rash	Rash evidence on skin	Text	C	Y/N
7.	Pulse	Pulse can be weak or normal	Text	C	Y/N
8.	B.P	Blood pressure may be weak or normal	Text	C	Y/N
9.	Hemmorhgic Menifestation	Small blotches of redness	Text	C	Y/N
10.	Tourniquet Test	Tourniquet test absent or positive	Num	N	10–40 Cont.
11.	WBC	Leucopenia/Normal	Num	N	0-10000 Cont.
12.	Platelette	Normal Platelet count Thrombocytopenia	Num	N	0-150000 Cont.
13.	Heamatocrit Test	Blood test that measures the percentage of red blood cells	Num	N	20–40 Cont.
14.	Rise in Hct	Rise value of % of Red blood cells	Num	N	20–45 Cont.
15.	IgG	Measure the level of immunoglobulin G/	Text	C	P/N
16.	IgM	Positive/Negative	Text	C	P/N
17.	Chest X-ray	For pleural effusion	Text	C	P/N
18.	UltraSound	Evidence of Ascites/size of liver & spleen	Text	C	P/N
19.	Bleeding	Spontaneous bleeding	Text	C	Y/N
20.	Classification	Dengue classification types.	Text	C	DF, DHF1, DHF2, DHF3

Pre-processing

The dataset used in this study was collected from various hospitals in Lahore[16], Pakistan, and includes information on the symptoms and diagnosis of actual dengue patients. The dataset consists of 376 instances, with 101 instances of dengue fever, 73 instances of DHF, 97 instances of DHF1, and 99 instances of DHF3. It includes 20 attributes, some of which are numerical, some categorical, and one class or target parameter. After preprocessing, we selected 15 relevant attributes for the analysis and converted them into binary values of 0s and 1s, except for the Tourniquet Test, WBC, and Platelet attributes which were converted into categorical values based on a set of ranges. The final dataset was split into two parts: a training dataset and a testing dataset. The training dataset was used to build the models, and the testing dataset was used to evaluate the models' accuracy. We used DT, RF, and NB algorithms individually and in combination, performing 10 iterations and varying the ratio of training and testing data. The final results showed improved accuracy when using the combination of all three algorithms.

The final experiment was conducted by running all three algorithms on the same dataset for each iteration, and then aggregating the results to determine the overall outcome. The combined results were found to be superior in all combinations of training and testing data ratios compared to individual results from each iteration. The model is designed to be adaptable and easy to improve upon in the future, as new factors can be added to enhance its performance and efficiency. The proposed predictive process of PDS is shown in Fig. 1 through a process flow design.

The final dataset used in this study consists of 400 records and 15 attributes. We divided the dataset into two parts: training data and testing data. The training data was used to train the models, while the test data was used to validate the models' outcomes. We ran each of the DT, RF, and NB algorithms individually, varying the ratio of training to test data and performing 10 iterations. After each iteration, we measured the accuracy of the models by passing the test data as input. Lastly, we combined the results of all three algorithms to obtain an aggregated outcome. The combined results showed improved performance compared to the individual results of each iteration. The models were optimized and flexible for future enhancements and adjustments. New factors could be easily added to the system to process dynamically and improve the effectiveness and efficiency of the system.

The study aims to classify and predict dengue outbreaks using a dataset and a predictive model. The dataset is analyzed to determine the key parameters that are important for predicting dengue. The pulse rate and bleeding are found to be the most significant elements and their variance has a significant impact on doctors' decision making and patient health outcomes. These parameters were chosen as the focus of this study due to the significant impact they have on dengue diagnosis and treatment, as shown in Fig. 2. The results of this study highlight the importance of monitoring pulse rate, bleeding, and platelet count in dengue patients to help diagnose and treat the disease effectively. This information can be used to develop more accurate dengue diagnostic models that are essential for controlling and preventing dengue outbreaks in the future.

The decision tree algorithm used to classify and predict the type of dengue fever patients are suffering from using a dataset. By analyzing the dataset, the study found that pulse rate, bleeding, and platelet count were the most significant parameters in predicting dengue fever. The Fig. 3, decision tree generated a tree plot that helped to understand the relationship between different parameters and the dengue fever type. The tree plot showed that a patient's pulse, bleeding, chest, and ultrasound results were important in determining the type of dengue fever. The decision tree algorithm is known for its interpretability, making it easy to understand the relationship between different parameters and the dengue fever type and to detect any errors.

The study conducted multiple iterations of RF with varying number of trees generated and the Fig. 4 illustrated results that as the number of trees increased, the out of bag error decreased. It was observed that when the number of trees surpassed 50, the error rate became consistent, remaining between 0.01% and 0.05%. The highest out of bag error observed during the experiment was 0.28%. The study concludes that random forest is an efficient algorithm for dengue classification with an accuracy of over 99%.

The following Table 2, 3, and 4, compare the performance of three different algorithms (Decision Tree, Random Forest, and Naive Bayes) in classifying and predicting dengue fever types. The performance of the algorithms was measured using a Confusion matrix, which shows the number of instances that have been correctly or incorrectly classified.

The Decision Tree algorithm had the highest accuracy at 99%, with 100% precision and recall for class 0, 99% precision and 96% recall for class 1, 97% precision and 100% recall for class 2, and 100% precision and recall for class 3.

The Random Forest algorithm had a slightly lower accuracy at 99.75%, with 100% precision and recall for classes 0, 1, and 3, and 96% precision and recall for class 2.

The Naive Bayes algorithm had the lowest accuracy at 69.25%, with 97.9% precision and 94% recall for class 0, 96.3% precision and 35.6% recall for class 1, 65.78% precision and 100% recall for class 2, and 100% precision and 98.95% recall for class 3.

Table 2

Decision Tree Confusion Matrix
	0	1	2	3	Precision
0	100	0	0	0	100%
1	1	72	0	0	99%
2	0	3	95	0	97%
3	0	0	0	98	100%
Recall	99%	96%	100%	100%

Table 3

Random Forest Confusion matrix
	0	1	2	3	Classification Error	Precision
0	100	0	0	0	0	100
1	0	73	0	0	0	100
2	0	3	95	0	0.031	96
3	0	0	0	98	0	100
Recall	100%	100%	96%	100%

Table 4

Confusion Matrix for the Estimation Sample Using NB
	0	1	2	3	Precision
0	94	2	0	0	97.90%
1	0	26	0	1	96.30%
2	6	45	98	3	65.78%
3	0	0	0	94	100%
Recall	94%	35.60%	100%	98.95%

The individual performance of algorithms, suggests that the Decision Tree and Random Forest algorithms have the highest accuracy in classifying and predicting dengue fever types, while the Naive Bayes algorithm has a lower accuracy. Additionally, in terms of precision, Naive Bayes performed better in classifying dengue type 0, while Decision Tree performed better in classifying dengue type 0 and 3 in terms of recall.

Focusing the Table 5, shows the performance comparison of the Decision Tree (DT), Random Forest (RF), Naive Bayes (NB) and Boosted Combo Model. It has been obtained by running 10 iterations of each algorithm on the same dataset. The Accuracy percentage of each algorithm is presented as the values in the respective columns.

Table 5

Performance comparison of DT, RF, NB, combo Model
Iterations	Decision Tree	Random Forest	Naïve Bayes	Boosted Combo
1	82%	92%	81%	92%
2	96%	90%	86%	96%
3	85%	97%	87%	97%
4	89%	95%	80%	95%
5	97%	98%	75%	98%
6	95%	98%	77%	98%
7	87%	91%	76%	91%
8	93%	95%	83%	95%
9	97%	90%	83%	97%
10	89%	97%	78%	97%
Min	82%	90%	75%	91%
Max	97%	98%	87%	98%
AVG	91%	94.3%	80.6%	95.6%

Table 5 shows the performance comparison of four algorithms (Decision Tree, Random Forest, Naive Bayes, and Boosted Combo Model) on a dataset. The table shows the average accuracy of each algorithm, obtained by running 10 iterations of each algorithm on the same dataset. The results show that Random Forest has the highest average accuracy of 94.3%, followed by Decision Tree with 91% and Naive Bayes with 80.6%. The Boosted Combo Model, which combines all three algorithms, has the highest average accuracy of 95.6%. The table also shows the minimum and maximum accuracy of each algorithm, which gives an idea of the range of values that each algorithm can achieve. The results suggest that Random Forest and Boosted Combo Model are the best performing algorithms, however, it's important to consider the relatively small sample size when interpreting the results. Further study should be made with a larger sample size to increase the robustness and generalizability of the results.

This paper describes a presumptive diagnosis system (PDS) that uses data mining techniques and decision tree, random forest, and naive bayes algorithms to aid in the early detection of dengue fever in areas with limited resources. The PDS achieved accuracy rates of over 90% in identifying dengue fever types, and a combination of these algorithms in the form of a boosted model produced the highest overall accuracy. The proposed PDS has potential to benefit underdeveloped countries where resources for diagnosis and treatment of dengue fever are limited and could also serve as a learning tool for new staff members in the healthcare sector. However, the study is limited by the dataset used and it is not a replacement for proper diagnosis and treatment by a qualified healthcare professional. Future work could involve incorporating more data points and attributes to increase accuracy, using machine learning and AI techniques to improve performance, evaluating the system's diagnostic accuracy in different contexts and populations, and testing the system's performance with other diseases.

Acknowledge

This to acknowledge that authors have no conflict of interest. Also, no organization, funding agencies and government not supported in any kind or cash.

Author contribution

Jibran Rasheed Khan researched and wrote the manuscript. Sehan Ahmed Farooqui, acquired dataset, build model, simulated and evaluated naïve basin and boosted algorithm. Syed Kawish Raza performed preprocessing, tested and evaluated decision tree, random forest algorithm. Farhan Ahmed Siddiqui supervised work, proofreading, edited script, while all authors reviewed the manuscript.

Funding

This work was not supported by any government, funding agencies in any kind or cash.

Data Availability

The hospital data used in work did not give written consent to share publicly or with anyone. Data only provided with consent to perform research work, due to the sensitive nature. Therefore, the data is not available.

Ethical approval of data taken from the authors of [16]. No personal information was requested.

Competing interests

The authors have no competing interests that might be perceived to influence the results or discussion reported in this paper.

WHO, “Dengue - Pakistan,” 2022. https://www.who.int/emergencies/disease-outbreak-news/item/2022-DON414 (accessed Nov. 12, 2022).
WHO, “Malaria - Pakistan,” 2022. https://www.who.int/emergencies/disease-outbreak-news/item/2022-DON413 (accessed Nov. 12, 2022).
WHO, “Dengue and severe dengue,” 2022. https://www.who.int/news-room/fact-sheets/detail/dengue-and-severe-dengue (accessed Nov. 12, 2022).
T. Arshad, A. Wajahat, A. Jabeen, and S. H. Ali, “Malaria and dengue outbreaks during a national disaster in Pakistan: A rising concern for public health,” J. Glob. Health, vol. 12, p. 03076, 2022, doi: 10.7189/jogh.12.03076.
Dailytimes, “Pakistan to record 2.7mn malaria cases by January says WHO,” 2022. https://dailytimes.com.pk/1009721/pakistan-to-record-2-7mn-malaria-cases-by-january-says-who/ (accessed Nov. 12, 2022).
G. Parthiban, “Applying Machine Learning Methods in Diagnosing Heart Disease for Diabetic Patients,” Int. J. Appl. Inf. Syst., vol. 3, no. 7, pp. 25–30, 2012, Accessed: Nov. 07, 2016. [Online]. Available: http://research.ijais.org/volume3/number7/ijais12-450593.pdf
A. Lewicki, K. Pancerz, and L. Puzio, “the Temporal Inference With the Use of Ant-Based Clustering Algorithm and Flow Graphs in the Problem of Prognosing Complications of Medical Surgical Procedures,” Hum. Technol., vol. 17, no. 3, pp. 213–234, 2021, doi: 10.14254/1795-6889.2021.17-3.3.
P. Mohapatra, N. K. Tripathi, I. Pal, and S. Shrestha, “Determining suitable machine learning classifier technique for prediction of malaria incidents attributed to climate of Odisha,” Int. J. Environ. Health Res., vol. 32, no. 8, pp. 1716–1732, 2022, doi: 10.1080/09603123.2021.1905782.
H. Gong, C. Xie, C. Yu, N. Sun, H. Lu, and Y. Xie, “Psychosocial factors predict the level of substance craving of people with drug addiction: A machine learning approach,” Int. J. Environ. Res. Public Health, vol. 18, no. 22, 2021, doi: 10.3390/ijerph182212175.
A. Dutta et al., “Early Prediction of Diabetes Using an Ensemble of Machine Learning Models,” Int. J. Environ. Res. Public Health, vol. 19, no. 19, pp. 1–25, 2022, doi: 10.3390/ijerph191912378.
K. Attai, Y. Amannejad, M. Vahdat Pour, O. Obot, and F. M. Uzoka, “A Systematic Review of Applications of Machine Learning and Other Soft Computing Techniques for the Diagnosis of Tropical Diseases,” Trop. Med. Infect. Dis., vol. 7, no. 12, 2022, doi: 10.3390/tropicalmed7120398.
F. Paquin, J. Rivnay, A. Salleo, N. Stingelin, and C. Silva, “Multi-phase semicrystalline microstructures drive exciton dissociation in neat plastic semiconductors,” J. Mater. Chem. C, vol. 3, pp. 10715–10722, 2015, doi: 10.1039/b000000x.
E. C. Montiel-Vázquez, J. A. Ramírez Uresti, and O. Loyola-González, “An Explainable Artificial Intelligence Approach for Detecting Empathy in Textual Communication,” Appl. Sci., vol. 12, no. 19, 2022, doi: 10.3390/app12199407.
M. Gobalakrishnan and D. Saravanan, “Antimicrobial activity of gloriosa superba, cyperus rotundus and pithecellobium dulce with different solvents,” Fibres Text. East. Eur., vol. 28, no. 4, pp. 67–71, 2020, doi: 10.5604/01.3001.0014.0937.
T. Sharma and M. Shah, “A comprehensive review of machine learning techniques on diabetes detection,” Vis. Comput. Ind. Biomed. Art, vol. 4, no. 1, 2021, doi: 10.1186/s42492-021-00097-7.
W. Farooqi and S. Ali, “A Critical Study of Selected Classification Algorithms for Dengue Fever and Dengue Hemorrhagic Fever,” in 11th International Conference on Frontiers of Information Technology, Dec. 2014, pp. 140–145. doi: 10.1109/FIT.2013.33.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this older preprint version

Read the latest preprint version →

Development and Evaluation of a Predictive Diagnostic System for Dengue Fever using Machine Learning Techniques

Status:

Version 1

Abstract

Figures

Introduction

Ict In Medicine

Material & Method

Machine Learning Algorithms

DATASET

Pre-processing

Research & Discussion

Conclusion

Declarations

References

Additional Declarations

Status:

Version 1