Multi-criteria decision analysis method for differential diagnosis of tropical febrile diseases

Abstract

This paper employs the Analytical Hierarchy Process (AHP) to enhance the accuracy of differential diagnosis for febrile diseases, particularly prevalent in tropical regions where misdiagnosis may have severe consequences. The migration of health workers from developing countries has resulted in frontline health workers (FHWs) using inadequate protocols for the diagnosis of complex health conditions. The study introduces an innovative AHP-based Medical Decision Support System (MDSS) incorporating disease risk factors derived from physicians’ experiential knowledge to address this challenge. The system’s aggregate diagnostic factor index determines the likelihood of febrile illnesses. Compared to existing literature, AHP models with risk factors demonstrate superior prediction accuracy, closely aligning with physicians’ suspected diagnoses. The model’s accuracy ranges from 85.4% to 96.9% for various diseases, surpassing physicians’ predictions for Lassa, Dengue, and Yellow Fevers. The MDSS is recommended for use by FHWs in communities lacking medical experts, facilitating timely and precise diagnoses, efficient application of diagnostic test kits, and reducing overhead expenses for administrators.

Introduction

Decision-making is a systematic process that helps people choose the best option from numerous alternatives and various organizations rely on decision support systems (DSSs) to improve decision-making activities. Multi-criteria decision analysis (MCDA) techniques are imperative tools utilized in developing DSSs, focused on aiding policymakers to choose the best alternative that achieves a given goal based on considered criteria.^1–3 According to Eltarabishi et al., ⁴ the ability to break down complex problems into simplest forms makes these methods the most recognized tools for handling decision-making problems. MCDA methods aid decision-making by considering multiple criteria, ⁵ delineating conditions, their importance, and influence.^6,7 Abdullah et al. ⁸ defines these methods as techniques for evaluating and selecting alternatives based on various criteria. In healthcare, selecting an appropriate MCDA method is crucial for intelligent systems aiding medical decisions.^9,10 Among some well-known MCDA methods such as Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS), VlseKriterijumskaOptimizacijaIKompromisnoResenje (VIKOR), Analytical Hierarchy Process (AHP), Analytic Network Process (ANP); Simple Multi-attribute Rating Technique (SMART), Measuring Attractiveness by a Categorical Based Evaluation (MACBETH), ELimination Et ChoixTraduisant la REalitè (ELECTRE), Preference Ranking Organization Method for Enrichment Evaluations (PROMETHEE), Evaluation of Mixed Data (EVAMIX), Evidence and Value: Impact on DEcisionMaking (EVIDEM), Multi-Attribute Utility Theory (MAUT) and AHPSort, ¹¹ the AHP was chosen in this study due to its effectiveness and suitability for group decision-making,^12–14 providing structured, comprehensive assessments of decision variables’ importance.^15–17

The incessant migration of healthcare workers from developing countries, such as Nigeria, to developed countries, has led to a severe dearth of health workers with antecedent negative consequences. This shortage of physicians has resulted in the training and employment of non-physician practitioners called community health workers or Frontline Health Workers (FHWs) for basic medical care delivery. FHWs in Nigeria use National Standing Orders, a written protocol as a guide to diagnose and treat common diseases at the primary care level. The works in Refs. 18 and 19 define a standing order (SO) as a procedure for identification of diseases as well as relevant treatments in the absence of a doctor by an authorized non-physician health worker. In multi-symptom, multi-disease situations, these FHWs lack adequate resources to carry out timely and accurate diagnosis with accomplishment of desired efficiency. Thus, the diagnostic process must be optimized to handle these multi-symptom, multi-disease conditions, especially for febrile illnesses.

Most tropical febrile diseases are confusable as they share a common manifestation or symptoms, making it difficult for physicians and FHWs to effectively diagnose. ²⁰ The management of high-burden and confusable febrile diseases requires an intelligent medical DSS (MDSS) as a tool for early detection and precise diagnosis of a patient’s health condition. This could lengthen the life span and enhance the well-being of the impacted population or group. This study develops an AHP-based multi-disease, multi-symptom tool for early and accurate differential diagnosis of febrile diseases by FHWs to increase healthcare access for people living in rural and resource-poor communities. Since it is challenging for physicians to decide on an illness based on the symptoms, our MDSS recommends the efficient utilization of Rapid Diagnostic Test (RDT) kits for diagnosing diseases. The febrile diseases considered in this study are some of the common diseases in Nigeria such as malaria, lower and upper respiratory tract infections, tuberculosis, acquired immunodeficiency syndrome (AIDS) and human immunodeficiency virus (HIV), enteric fever, lower and upper urinary tract infections, lassa fever, dengue fever and yellow fever. ²¹ Anopheles mosquitoes transmit plasmodium parasites, causing malaria with symptoms like fever, headache, and fatigue. Antimalarial drugs and antimalarial chemoprophylaxis are some pharmaceutical interventions that combat malaria,^22,23 while interventions like indoor spraying and insecticide-treated bed nets are non-pharmaceutical. ²⁴ Enteric fever caused by Salmonella Typhi, leads to severe symptoms like abdominal pain and can be fatal if untreated. Treatments include antibiotics, vaccines, and hygiene improvements.^25,26 HIV weakens the immune system, leading to opportunistic infections. Antiretroviral therapy and safe sex education help manage HIV.^27,28 Respiratory tract infections (RTIs) affect the upper and lower airways, with symptoms such as cough and chest pain. RTIs are caused by bacteria, viruses, and fungi. Treatments include antibiotics, antiviral drugs, vaccination, and respiratory hygiene.^29–31 Urinary tract infections (UTIs) result from bacterial entry into the urethra, causing symptoms like abdominal pain and frequent urination. The works in Refs. 32–34 report that antibiotics and personal hygiene practices are good treatment options for UTIs.

This study considers risk factors contributing to tropical diseases, acknowledging their complexity and shared risk factors as in Uzoka et al. ³⁵ The MDSS will aid physicians, especially in remote areas, but its effectiveness can be boosted with user-friendly software and mobile health app considerations for data security. No previous research has examined diagnosing and treating eleven febrile diseases using an MDSS. The following is a list of our specific contributions to this paper. (a) gathering data from patients and physicians to create datasets; (b) employing data analytics to construct a multi-disease, multi-symptom AHP model; (c) devising a linear mathematical model, aggregate diagnostic factor index (ADFI), for disease assessment triggered by AHP; (d) developing an AHP-based MDSS for differential diagnosis; (e) setting medical expert thresholds for disease prediction; (f) designing a spreadsheet dashboard for displaying results; (g) assessing the MDSS’s performance against laboratory tests for diagnostic accuracy, reliability, and effectiveness.

The remaining content is arranged as follows: A Section on Related Works examines cutting-edge methods for creating MDSSs for illness diagnosis and therapy. The experimental design of the proposed AHP-based MDSS is presented in the methodology section using field datasets from physicians’ experience, and the discussion of results section discusses the results and their implications after comparing them with the diagnoses of medical experts and validated laboratory test results. The article is concluded with suggestions for more research.

Related works

MCDA methods in healthcare

Healthcare demands swift, accurate decision-making with less error. ⁴ Analyzing interventions and MCDA methods highlights their importance in healthcare problem-solving. Reviews on MCDA in healthcare often focus on supporting investment decisions, health technology assessment (HTA), coverage determinations, authorization, medical automation, performance measurement, disease management, research funding allocation, and more. ³⁶ Alharbi et al. ¹ evaluated healthcare information systems’ efficiency using a combined TOPSIS and Fuzzy AHP framework, identifying access control and software security as vital factors. Ayan et al. ³⁷ discussed various weighting techniques for MCDA tasks in healthcare, including entropy, AHP, Level Based Weight Assessment (LBWA), CRiteria Importance Through Intercriteria Correlation (CRITIC), Full Consistency Method (FUCOM), Criterion Impact Loss (CILOS), Best Worst Method (BWM), and Method based on the Removal Effects of Criteria (MEREC). These techniques can be divided into three groups: objective, subjective, and hybrid. Hybrid techniques blend subjective and objective approaches, while subjective methods rely on decision-maker-assigned weights and objective methods employ mathematical algorithms. Selecting an appropriate method poses challenges, with objective methods offering reproducible results and greater accuracy, particularly in complex decision scenarios with numerous criteria. They are recommended for situations where expert input is difficult to obtain.

The BWM is a decision-making technique focused on identifying the best and worst criteria within a set. In Refs. 38, BWM was observed to be suitability for health decisions due to its adaptability and consistent outcomes, suggesting its growing preference in health applications. However, Pamučar et al. ³⁹ pointed its limitations in scenarios with multiple best or worst criteria. The FUCOM algorithm, requiring fewer pairwise comparisons, is used in various healthcare evaluations, such as healthcare waste treatment selection and pandemic resilience.^40–42 Despite its efficiency, FUCOM’s fewer comparisons may hinder conveying subtle preference variations, potentially leading to less accurate decisions. Similarly, the LBWA, another subjective-based method, ranks criteria by relative importance, offering a solution for decision-makers struggling with assigning precise weights. Although LBWA is relatively new, it finds application in healthcare decision-making, albeit with fewer comparisons, akin to FUCOM.^39,43,44 Objective-based methods such as AHP, entropy, CRITIC, CILOS, and MEREC are employed in decision-making. CRITIC identifies criteria contrasts using correlation coefficients and standard deviations. CILOS adjusts criteria significance as they approach ideal values, overcoming entropy method’s limitations. According to Ref. 45, MEREC calculates weights based on criteria’s removal effects on alternatives, aiding decision-makers in excluding specific criteria. Unlike other methods, MEREC focuses on exclusion rather than inclusion principle. ⁴⁶ AHP constructs comparison matrices to determine criteria weights, aggregating eigenvectors to gauge alternatives’ relative importance.

Although this paper focuses on disease diagnosis and treatment, Table 1 shows some MCDA methods and their applications in healthcare.

Table 1.

MCDA methods in healthcare.

Reference	MCDA method	Objective	Result
Goetghebeur et al 47	Evidence and value impact on decision making (EVIDEM()	Bridging HTA and efficient healthcare decision making	Improvement in efficacy and effectiveness is most important to decisions; the framework facilitates data reusability, re-evaluation of interventions as new data are generated
Cox et al 48	MACBETH	Determining which animal and human infectious diseases are most likely to be impacted by climate change on a global scale	The diseases that were found to rank highest using the weighting method were Chagas, Giardiasis, and West Nile virus, while coccidioidomycosis was ranked lowest
Amaral and Costa 49	PROMETHEE II	Enhancing resource allocation and decision-making in an emergency department	The ranking revealed that, out of all the six criteria taken into consideration, alternative A1 had the highest score, making it the best option available to improve patient flow
Wahlster et al 50	EVIDEM V2.2	Examining the opinions and preferences of German stakeholders when evaluating a heart-pulmonary sensor as a healthcare intervention	Policy makers reported the lowest value of 0.40 for the sensor, compared to 0.48 for all participants, and industry representatives reported the highest value estimate of 0.68 (on a scale of 0 to 1). Participants said that the majority of qualitative factors ought to be taken into account and their influence on the quantitative evaluation openly documented
Ghandour et al 51	MCDA	Examining whether it is possible to prioritize cardiovascular disease (CVD) policies in Palestine, Syria, Tunisia, and Turkey by working with policymakers and stakeholders to apply the MCDA method	A list of policies specific to each of the four countries was created by rating 25 different policies related to CVD prevention and control. The majority of the most important policies were those at the level of the health system and the population
Uzoka et al 16	AHP	Creating an AHP model for a multi-model diagnosis of seven tropical illnesses that are difficult to diagnose: Measles, chicken pox, urinary tract infection, yellow fever, malaria, typhoid, and hepatitis B	The findings indicate that fever was a prevalent and significant symptom in each of the seven tropical diseases
Angelis and Kanavos 17	Multi-attribute value theory	Evaluating the value of new medicines in HTA	The advanced value framework can be used by decision makers and stakeholders to measure the value of new medicines. It considers deployment in real-life taking note of budget constraints to achieve the best value-for-money
Alsalem et al 52	VIKOR	Proposing a DSS based on MCDM for benchmarking and assessment to choose appropriate multiclass classification models for acute leukemia	A significant gap was found that prevented the evaluation and benchmarking process from meeting all of the crucial requirements for the detection and classification
Jiménez et al 53	EVIDEM framework	Evaluating the worth of selexipag, a novel oral prostacyclin receptor agonist, in treating pulmonary arterial hypertension (PAH) in Spain using the reflective MCDA approach	MCDA offers a transparent and standardized approach to help with decision-making when evaluating various criteria related to selexipag’s overall value contribution to PAH treatment. Oral selexipag was thought to be a more beneficial treatment for PAH than iloprost
Berumen et al 54	EVIDEM	Facilitating effective resource allocation, improving access to medical devices, and creating ready-to-use guidelines for the selection of devices to establish, maintain, and operate necessary clinical units within the continuum of care for a defined set of cancer types	Capital equipment (82 medical equipment, 62 laboratory and pathology equipment, 28 quality assurance devices, and 416 surgical instruments, including 52 single instruments and 28 sets, kits, or trays containing 364 surgical instruments) are among the medical devices. Other items include single-use equipment, consumables, reagents, accessories, and protection devices (26 for laboratory and pathology and 22 for personal protection); 23 radiation protection and monitoring devices; 180 single-use and disposable medical supplies; 108 solutions and reagents; 70 glassware items and utensils; and three software or information systems
Babashahi 55	Potentially all pairwise rankings of all possible alternatives (PAPRIKA)	Funding for health research should be prioritized, including non-communicable diseases	The findings of the comparison between the AHP in expert Choice software and the PAPRIKA in 1000minds show that both programs employ distinct theoretical underpinnings at various points in the decision-making process, from gathering participant preferences to computing alternative scores and criterion weights. To elicit participant preferences, AHP uses ratio scale-based paired comparison questions, while PAPRIKA uses choice-based pair-wise comparison questions
Schey et al 56	EVIDEM	Developing an MCDA framework in HTA for resource allocation and aggregate orphan drug values based on attributes	Determined the five (5) characteristics that are most commonly taken into consideration: Tolerability, quality of evidence and safety, comparative effectiveness or efficacy, unmet need (or availability of therapeutic alternatives), and severity of the disease
Ferreira-Coimbra et al 57	Weighted criteria	Identifying diagnostic tools for Ventilator-associated pneumonia (VAP) diagnosis in clinical practice	Performed a variety of diagnostic methods, including biomarkers, images, and the culture of clinical specimens, to diagnose ventilator-associated pneumonia; Gram staining was found to be the most effective method
Uzoka et al 58	AHP	Creating a diagnostic system for typhoid fever based on the AHP model	The system yields 78.91% correct diagnosis when compared with the patient dataset
Albahri et al 59	AHP and group VIKOR	Describing a multi-biological laboratory investigation structure that uses integrated AHP and group VIKOR methods to prioritize COVID-19 patients	The proposed framework could adopt and extend the criteria or change the subjective judgment of medical experts
Albahri et al 60	Integrated Entropy–TOPSIS	Describing a comprehensive framework for making decisions that will help prioritize COVID-19 patients and determine the health status of carriers who do not exhibit any symptoms	The suggested framework can classify medical conditions as mild, serious, or critical, prioritizing patients in line and taking asymptomatic carriers into account
Villanueva et al 61	Reflective MCDA	Determining what constitutes a good treatment and management strategy for patients with drug-resistant epilepsy (DRE) who experience focal-onset seizures (FOS)	The study employed reflective MCDA methodology to assess the efficacy of treatment and management of focal-onset seizures in patients with drug-resistant epilepsy
Andrés-Nogales et al 62	FinMHU-MCDA	Applying MCDA in the reimbursement of orphan medicinal products (OMP) drugs in Spain	Nine of the thirteen criteria that were chosen were thought to be pertinent for making decisions and were linked to a higher relative importance
Razzaq et al 63	F-AHP and F-VIKOR	Assisting decision-makers in determining a viable preventive strategy based on intervention efficacy and public acceptance	It was discovered that F-AHP and F-VIKOR worked quite well together to translate the real data
Mumtaz 64	AHP	Determining the obstacles that prevent senior citizens from using the e-healthcare system in order to better support the creation of a digital healthcare system for this demographic in the future	The results indicate that barriers related to health and ability rank highest, followed by barriers related to the socio-environment and attitudes
Razzaq et al 65	F-AHP and F-PROMETHEE	Providing awareness tool for researching a disease that has not been thoroughly studied yet in terms of its virulent transmissibility, symptoms, and preventative measures	The evaluation’s findings showed that posting brief videos or written content on social media networks was less expensive and more successful in gaining public acceptance
Lin et al 66	TOPSIS and CRITIC	Examining how an experimental society shares information by using a virtual messaging platform	The evaluation’s findings indicate that the criterion with the largest weight, extensive interaction, is 0.1950, followed by the other criteria

MCDA methods for disease diagnosis and treatment

In under-developed countries like Nigeria, confusable tropical disease symptoms pose treatment challenges due to limited healthcare infrastructure.^67,68 Asuquo et al. ⁶⁹ observed there is continual increase in patients’ waiting times and hospital stays, especially in remote areas. Implementing a clinical diagnostic system can provide swift, accurate decision-making tool for nursing staff and less-experienced physicians, improving disease severity evaluation and prompt treatment, crucial where doctors and lab. tests are scarce. This is capable of reducing mortality, morbidity rates, and enhancing overall disease control.

Physicians’ experiential knowledge aids disease diagnosis but requires automated historical data processing due to conflicting symptoms. ¹⁶ Mathematical reasoning methods struggle with medical complexity by merely reporting disease presence or absence.^70,71 Hence, an intelligent MDSS is needed. This study presents an AHP-based MDSS for prompt and accurate differential diagnosis of 11 febrile diseases in Nigeria, addressing the challenges of medical decision-making in tropical regions. Various studies have focused on diagnosing tropical diseases, but most consider a limited number of diseases, such as malaria, yellow fever, and typhoid.^16,35,72 Some only develop single-disease models, identifying significant symptoms for each disease.^{14,58,73–76} While previous studies integrate physician judgment,^14,73,77 none has examined up to 11 diseases or compared DSS performance with confirmed diagnoses. This study, akin to Uzoka et al. ¹⁶ and Osamor et al., ⁷¹ presents a user-friendly interface for field health workers (FHWs) and outlines earlier research on DSS development methods. Table 2 provides an overview of previous research on cutting-edge methods for developing DSSs in the last decade for disease diagnosis and treatment capable of handling complexities due to uncertainty and imprecision in medical data.

Table 2.

Summary of previous studies on MCDA methods for disease diagnosis and treatment.

Reference	Objective	Tool	Sources of information	Diseases considered	Result
Ijzerman et al 78	Compare the effectiveness of AHP and conjoint analysis (CA) in getting patients’ preferences for different stroke rehabilitation treatment options	AHP and (CA)	Patients	Stroke	The two most popular treatments were orthopedic shoes (OS) and soft tissue surgery (STS); OS was most preferred when using CA weights and STS was most preferred when using AHP weights
Diaz-Ledezma and Parvizi 79	Analyze surgical approaches for cam FemoroacetabularImpingement (FAI)	AHP	Physicians	Cam femoroacetabular impingement	AHP model showed that mini-open anteriorapproach of the hip (MO) is the best surgical approach for cam FAI followed by hip arthroscopy (HA), and surgical hip dislocation (SHD)
Chung et al 80	Set performance goals high for colorectal cancer treatment to make the pay-for-performance (PFP) system easier to implement	AHP	Physicians	Colorectal cancer	The most significant measurements were the percentage of patients with colorectal cancer who had pre-operative exams that included abdominal ultrasound, computed tomography, or MRI, and the percentage of patients with stage I–III colorectal cancer who had a wide surgical resection that was recorded as having a “negative margin” and the percentage of patients with colorectal cancer who had surgery and received a pathology report that included details on tumor size and node differentiation
Diaz-Ledezma et al 81	Establish a systematic and repeatable method for identifying periprosthetic joint infections (PJI) in the hip and knee while maximizing available resources and minimizing the degree of unpredictability in medical professionals’ subjective assessments	AHP	Physicians	Periprosthetic joint infection	The best alternative to diagnose knee PJI in medicare patients was screening with serum markers (erythrocyte sedimentation rate (ESR)/C-reactive protein (CRP)) and then arthrocentesis in positive cases (strategy 1). This was followed by serum markers requested concurrently with arthrocentesis (strategy 3) and then immediate arthrocentesis (strategy 2). For the hip PJI model, the same ranking of alternatives was noted
Suner et al 82	Provide a web-based clinical DSS to help patients with rectal cancer choose treatments that may be helpful	AHP	Physicians	Rectal cancer	It was discovered that the consistency values for the most common cases were 100% for the second decision step and 80% for the first. Consistency for the first decision step was 50% and for the second decision step was 80% in rare cases
Maranate et al 83	Prioritize clinical risk factors of obstructive sleep apnea severity using fuzzy AHP	Fuzzy-AHP	Physicians and patients	Obstructive sleep apnea (OSA)	The findings show that nighttime symptoms ranked highest among clinical risk factors for severe, moderate, mild, and No OSA.
Xu et al 84	1) Determine which test characteristics are crucial for making decisions	AHP	Patients	Colon cancer	The most important factors were test accuracy, difficulties, and test preparation; a small percentage of patients chose FIT over colonoscopy; regarding accuracy, patients favored colonoscopy over FIT; however, when it came to preventing complications and getting ready for the test, they preferred FIT over colonoscopy
	2) Examine patient preferences for colonoscopy versus fecal immunochemical testing (FIT) for colorectal cancer screening
Mühlbacher et al 85	After suffering from acute coronary syndrome, ascertain the patients’ top priorities for long-term medication therapy (ACS)	AHP	Patients	Acute coronary syndrome	Patients demonstrated a strong sense of priority for the attribute “reduction of mortality risk” followed by “prevention of a new myocardial infarction”, followed by “side effects: Dyspnea and then bleeding”; while “frequency of intake” was the least significant feature
Samuel et al 86	Heart failure risk prediction	ANN and fuzzy AHP	Physicians and patients	Heart failure	When compared to the traditional ANN method, the suggested hybrid approach could achieve an average prediction accuracy of 91.10%, which is 4.40% higher
Agapova et al 87	Assess the relative importance of diagnostic imaging tests in diagnosing suspected appendicitis	AHP	Physicians	Appendicitis	The outcome demonstrated that the best test is contrast-enhanced computed tomography
Jiménez et al 53	Evaluate the potential benefits of selexipag, a novel oral prostacyclin receptor agonist, for the treatment of PAH in Spain	EVIDEM frame work	Physicians and patients	Pulmonary arterial hypertension	Using these quantitative criteria—comparative efficacy, patient-reported outcomes, preventive benefit, therapeutic benefit, other medical costs, and other non-medical costs—oral selexipag for PAH treatment was deemed to be a more valuable treatment than iloprost
Albahri et al 59	Outline a novel multi-biological laboratory examination framework that uses integrated MCDA methods to prioritize patients with COVID-19	AHP-VIKOR	Physicians and patients	COVID-19	The AHP–VIKOR method was successfully integrated to address patients with COVID-19’s prioritization issues in both individual and group decision-making (GDM) contexts
Sayan et al 88	Select the most effective diagnostic method for SARS-CoV-2 diagnostic tests	Fuzzy PROMETHEE and fuzzy TOPSIS	Physicians	COVID-19	The study’s results using both of the suggested MCDM methods showed that computerized tomography of the chest (chest CT) was the most successful method for diagnosing COVID-19
Fu et al 89	Propose a data-driven MCDA framework for diagnosing thyroid cancer	Data-driven MCDM framework	Physicians and patients	Thyroid cancer	Learning criterion weights and their constraints results in a 40.53% increase in diagnostic accuracy
Panattoni et al 90	Examine whether a time trade-off (TTO) or direct rank ordering can be used in an MCDA to evaluate the choice of chemotherapy in a sizable population-based sample	Direct rank ordering	Physicians and patients	Breast cancer	The age range of the survivors was 25 to 74 years, with 73.9% having stage I tumors; the TTO score was 94.2%, and the response rate for the rank order was 81.0%
Madeira et al 91	Assess the benefit-risk ratio on the efficacy and safety of all phosphodiesterase type 5 inhibitors (PDE5i) in men with erectile dysfunction	Network meta-analysis, the surface under the cumulative ranking analysis, and stochastic multi-criteria acceptability analyses (SMAA)	Patients	Erectile dysfunction	Avanafil and lodenafil were less successful interventions, while sildenafil 25 mg and 50 mg were the most enhanced/effective treatments compared to all other interventions. Taladafil 20 mg and 10 mg came in second and third, respectively. More side effects, particularly flushing and headaches, were brought on by the medication mirodenafil 150 mg. Whereas vardenafil and udenafil were more likely to induce nasal congestion, sildenafil 100 mg was more associated with visual problems
Karrer et al 92	Examine the opinions and preferences of medical professionals and patients about the current thyroid nodule interventions	EVIDEM	Physicians and patients	Thyroid cancer	The highest value contributor was ‘comparative effectiveness’. For the physicians’ group, ‘comparative safety’ was given higher value while the patients’ group has ‘type of preventive benefits’ contributing more positively to the value estimation
Calvet et al 93	Based on a national quality certification program in Spain, update the current set of quality indicators (QIs) for inflammatory bowel disease comprehensive care units (ICCUs) from a multi-stakeholder standpoint	AHP	Physicians and patients	Inflammatory bowel disease	The final set of 67 QIs was divided into three categories based on their relative importance: Structure (23 QIs), process (35 QIs), and outcome (9 QIs). The most crucial requirement for ICCUs was multidisciplinary management, which was followed by continuity of care, clinical care standardization, and incorporating patient-reported outcomes
Uzoka et al 58	Develop a DSS for diagnosis of typhoid fever	AHP	Physicians, patients	Typhoid	The system yields 78.91% correct diagnosis when compared with the patient dataset

Methodology

Study design and sampling method

The methodology deployed in this study for the task of febrile disease diagnosis is presented in Figure 1. Before formulating the AHP structure and developing its model, a field study in selected states in Nigeria’s Niger Delta region—Rivers, Cross River, Akwa Ibom, and Imo—resulted in the creation of a febrile disease dataset. Data were collected using two instruments validated by domain experts with the use of the Open Data Kit app. ⁹⁴ Sixty-two (62) doctors working in secondary and tertiary care facilities—both public and private—who specialize in diagnosing febrile illnesses provided their experiential knowledge to one data source. A patient’s consultation tool, intended to help doctors gather information about their symptoms and record initial diagnoses along with recommendations for further investigation, was the other source of data. The patients’ dataset was used as a baseline to test the performance of the AHP model developed from the physicians’ experiential knowledge to assess the degree of precision in the febrile disease differential diagnosis.

Figure 1.

Study methodology for febrile disease diagnosis.

The sections of the physician knowledge extraction tool are A, B, and C. Section A contains demographic information of physicians while Section B provides physicians the opportunity to share their knowledge on the causative relationships between each of the diseases and associated symptoms. The causative relationships were rated on a linguistic scale of 1-5 for none, low, moderate, high, and very high, respectively. The instrument also provided an opportunity for the physicians to indicate their level of confidence in the rating on a scale of 1-10, where a confidence level of 1 denotes the lowest and a level of 10 the highest. In this study, we recognized the possibility that genetics, lifestyle, and exposures could predispose individuals to some of these diseases. ³⁵ Section C gives doctors the chance to assign a numerical value (1 = No effect, 2 = Weak effect, 3 = Moderate effect, 4 = Strong effect, 5 = Very strong effect) to each risk factor for the disease under consideration, such as exposure to mosquito bites, genetic conditions, high cholesterol, high blood pressure, and travel to endemic regions. These risk factors were captured during patient consultation and became a component of the decision support filter of our diagnostic model. The identified symptoms became criteria while the diseases became alternatives. The study goal is to perform differential diagnosis of eleven febrile diseases with the possibility of uncovering co-morbidity.

Data analysis

Demographic profile of physicians

Descriptive statistics, including frequency, percentage, mean, and standard deviation, were utilized to conduct exploratory data analysis on the dataset of physicians. The demographic characteristics of the physicians are presented in Table 3, encompassing location, gender, age, and years of professional experience. Among the 62 sampled physicians, 69% were male, with the remaining being female contributors to the dataset. Age distribution revealed that the majority of physicians (51.61%) fell within the 31–40 age bracket, followed by 41–50 years (29.03%), and those over 50 years (12.9%), while 6.45% were aged 30 or below. Notably, 80.64% of physicians were between the ages of 31 and 64. Regarding experience, the analysis showed that 51.6% had ten or more years of experience in the field, whereas 9.7% had 5 years or less. This summary provides an overview of the physicians’ demographic characteristics and experience distribution within the study.

Table 3.

Location, age, gender, and experience distribution of physicians.

State	No. of doctors	Gender
		Male	Female
Akwa Ibom	16	12	4
Cross river	16	12	4
Imo	15	12	3
Rivers	15	7	8
Total	62	43	19

Age (years)	Frequency	Percentage
<=30	4	6.45
30<age<=40	32	51.61
40<age<=50	18	29.03
>50	8	12.9
Total	62	100

Work experience (years)	Frequency	Percentage
<=5	6	9.7
5<Yr<=10	24	38.7
10<Yr<=15	14	22.6
>15	18	29.0
TOTAL	62	100

Sampling distribution of the population mean of physicians’ assertions

For each febrile disease, a physician provides the interrelationship between the symptoms peculiar to the disease and the confidence level (CL) or score for each response. For example, if the interrelationship between a symptom (e.g. fever) and disease (e.g. malaria) according to a physician is 5 (very high), the physician should also pick a CL between 1 and 10 to corroborate that when a patient has malaria, the patient is certainly going to have a very high fever. Table 4 shows a few physicians’ responses to some symptoms of malaria and their scores. Only a few observed symptoms of malaria such as generalized body pain (GENBDYPN), Headache (HDACH), Fever (FVR), High-grade Fever (HGRDFVR), Chills and rigor (CHLNRIG), Bitter taste in the mouth (BITAIM), Fatigue (FTG), Joint pain (JONPN) and Prostration (PRTN) are shown in the table to maximize space.

Table 4.

Interrelationship connecting malaria symptoms and their confidence level.

S/N	GENBDYPN	HDACH	FVR	HGRDFVR	CHLNRIG	BITAIM	FTG	JONPN	PRTN
1	4	4	5	4	3	3	4	3	3
2	4	5	3	5	3	5	4	3	2
3	4	4	5	4	4	4	5	5	4
4	4	4	5	4	4	3	4	4	3
5	4	4	5	3	4	3	4	3	4
6	4	4	4	3	4	4	5	4	3
7	4	4	5	4	4	4	4	4	3
8	3	4	3	2	4	3	3	3	2
9	5	5	5	5	5	5	5	5	4
10	3	4	4	3	4	4	4	3	3
11	3	4	5	5	4	3	3	4	3
12	3	4	5	5	4	3	5	4	4
13	3	4	3	4	4	4	3	4	4
14	3	3	4	4	3	3	3	3	2
15	5	4	5	5	4	5	3	2	3
16	3	4	5	4	4	3	3	2	4
17	4	4	4	5	4	3	3	3	2
18	4	4	5	5	4	4	3	2	2
19	4	4	4	4	3	5	4	3	1
20	4	3	5	3	3	4	3	2	2

Medical researchers commonly employ parametric tests to accurately estimate data with greater statistical power. They rely on the Central Limit theorem (CLT) to assume the population’s probability distribution. CLT posits that with a sufficiently large sample size from a population of known variance, the sampled variables’ mean will approximate the population mean. Moreover, these samples tend to follow a normal distribution, with variances approaching that of the population as sample size increases. Mathematically, CTL states that, if $X_1,X_2,X_3,\dots ,X_n$ are random samples from a population with mean, $\mu$ and finite variance, ${\sigma }^2$ and if ${\overline{X}}_n$ is the sample mean of the first $n$ samples, then the limiting form of the distribution, given in equation (1), is a standard normal distribution with sample standard deviation, ${\sigma }_{\overline{x}}=\frac{\sigma }{\sqrt{n}}$ .^95–96

$Z=\underset{n\to \infty }{\lim }\left(\frac{{\overline{X}}_{n-}\mu }{{\sigma }_{\overline{x}}}\right)$

(1)

We applied standard score (Z score) and CLT on the data in rating the causative relationship between each disease and its associated symptoms for further analysis. The mean and standard deviation of the physicians’ assertions were generated from the CLT. According to Andrade, ⁹⁷ the Z score informs the researcher how far below or above a particular score is from the mean. Microsoft Excel 2016 spreadsheet was the tool used for statistical analyses in this work with the help of RAND and RANDBETWEEN functions for generating 30 random samples 30 times from the 62 physician population for all 11 febrile diseases, resulting in a sample size of 900 for each disease. The RAND function returns uniformly distributed random numbers every time the worksheet is refreshed. The RANDBETWEEN function returns random numbers between a specified range. The computed sample mean for each of these samples produces a sampling distribution and according to CLT, the sampling distribution appears more and more like a normal distribution as the sample size increases. This information was used to calculate the population mean required in the AHP methodology for decision-making analysis. The mean opinion of physicians with regards to the impact of each symptom on the disease on a five-point scale (none; low; moderate; high; very high) was converted or normalized to Saaty’s nine-point scale ⁹⁸ (see Table 5) using equation (2), as follows:

$\text{Normalized}\hspace{0.17em}\text{value}=\frac{Mean\hspace{0.17em}Opinion}{5}x\hspace{0.17em}9$

(2)

Table 5.

Saaty preference scale.

Saaty scale	Description
9	Far better than the threshold
8	Much better than the threshold
7	Better than the threshold
6	Slightly better than the threshold
5	About the same as the threshold
4	Slightly worse than the threshold
3	Worse than the threshold
2	Much worse than the threshold
1	Far worse than the threshold

AHP modeling for febrile disease diagnosis

Conceptual disease-symptoms modeling

An exploratory analysis of the physicians’ dataset reveals a total of forty-three (43) symptoms and eleven (11) diseases (see Table A.1), out of which nausea and vomiting (NUSNVMT), nausea (NUS), vomiting (VMT) have been changed to NUS-VMT; reducing the symptoms to 41 in all. So, a patient with a symptom of nausea or vomiting or both is represented as NUS-VMT. From our analysis, twenty (20) symptoms were interrelated with two or more febrile diseases, presented in Table A.2a as a multi-disease-symptom model. Table A.2a shows that while FVR is a symptom related with all the 11 diseases; FTG and HGDFVR were related with 8 of the diseases; NUS-VMT and LWGDFVR each with 5 of the diseases; CGH, CHSPN, HDACH, SRTRT, and CHLNRIG each with four of the diseases; BLDFASG, DIFBRT, and JNTNMSCPN each related with three of the diseases; BLDYURN, CLYDURN, GENBDYPN, NGTSWT, WGTLS, WHZ, and ADPN each with 2 of the diseases. However, the single disease-symptom model in Table A.2b indicates the symptom that contributes to triggering the diagnosis of a specific disease. For instance, for a patient with a symptom of YELOEY, the YELFVR disease model will be triggered for diagnosis results and treatment while TB disease will be triggered for LMP symptom, and so on. A combination of the two disease-symptoms model further indicates that a patient with symptoms of ADPN, FTG, FVR, GENBDYPN, HDACH, JONPN, BITAIM, PRTN, and/or HGDFVR will cause the MAL disease model to be triggered while symptoms like CGH, FVR, HGDFVR, LWGDFVR, SRTRT, CTRH, FOLBRT, and/or EPNOTG will trigger the URTI disease model.

Intelligent disease diagnosis system with AHP model

The architectural design of the MDSS for febrile disease diagnosis is shown in Figure 2, with medical experts, a knowledge base, a diagnostic system, decision support filters, an AHP diagnostic engine, a patient, FHW, and a user interface as major components. The user interface enables a FHW to input appropriate information obtained from interaction with a patient such as symptoms and risk factors into the diagnostic system. Structured experiential knowledge of the study domain obtained from medical experts is represented in the knowledge base. The patient examination and test results, along with the risk factors that make a patient more susceptible to these febrile diseases, are adequately included in the knowledge base. The decision support filters in the diagnostic system use this data to mimic the actions of a skilled physician utilizing the AHP diagnostic engine to initiate the relevant individual febrile disease models as a predicted result for additional examination. The AHP diagnostic engines help identify suspected illnesses based on patient symptoms and risk factors, allowing doctors to recommend the best course of action. They also differentiate between the symptoms of feverish illnesses. The FHW can as well visualize the system’s predicted diagnosis on the user interface for prompt and precise healthcare decision-making, medical advice, and treatment.

Figure 2.

System architecture of the AHP-based MDSS for febrile disease diagnosis.

Due to the challenges in large-scale group decision-making, we develop explicit, linear consensus models stored in the AHP knowledge base to trigger each disease for intelligent diagnosis based on the contribution of its associated symptoms from physicians’ opinions and preferences. Thus, a threshold is set for each disease prediction by the AHP diagnostic engine and for the possibility of co-morbidity where there are overlapping symptoms. The linear models provide a formula for computing the probability of occurrence of any of the eleven diseases based on the linguistic value of symptoms: 1 = absent; 2 = very low; 3 = low; 4 = moderate; 5 = high; 6 = very-high.

The symptom-by-symptom pairwise comparison for each disease alternative was evaluated using the comparison matrix $A$ in equation (3), with each leading diagonal equal to 1, i.e. $a_{ij}=1,\forall i=j$ . The comparison matrix is formed by allocating the value of pairwise comparison of the element in the row $i$ with the element in column $j$ into the position $a_{ij}$ . A high value shows greater importance in the element comparison and vice versa. In this case, $n$ is the number of symptoms to be evaluated, $C_i$ is the $i_{th}$ symptom, and $a_{ij}$ is the comparison of the $i$ symptom with respect to the $j$ symptom. The AHP technique then ascertains the relative weight of each symptom and combines the eigenvectors for every comparison matrix until the compound weight for each disease is attained. The values of the final weight vector show the relative importance of every disease with regard to the goal. A policymaker then utilizes the eigenvectors to rank the diseases for suitable decision-making. The normalized vector matrix is obtained using equation (4), where $a_{ji}$ represents the element at row $j$ and column $i$ of the respective symptom versus symptom comparison matrix. Finally, the relative weight or weighted eigenvector, $W_i$ is computed using equation (5) as the row average of the resulting normalized matrix such that ${\displaystyle \sum }_{i=1}^n{W}_i=1.$ For consistency analysis, equations (6) and (7) were used to estimate the consistency index (CI) and consistency ratio (CR) of each performance matrix, using Saaty’s random consistency index (RI) estimates in Table 6, where n is the number of symptoms. If $CR\le 0.1$ , the level of inconsistency is acceptable, else the inconsistency is high and the decision-maker is advised to revise the elements of $a_{ij}$ to realize a more consistent matrix.

$A=\begin{array}{l}\begin{array}{l}{C}_1\\ C_2\\ C_{3\hspace{0.17em}}\end{array}\\ \dots \\ C_n\end{array}\left[\begin{array}{l}\begin{array}{lll}{C}_1& C_2& \begin{array}{ll}{C}_{3\hspace{0.17em}}& \begin{array}{ll}\dots & C_n\end{array}\end{array}\\ 1& a_{12}& \begin{array}{ll}{a}_{13\hspace{0.17em}}& \begin{array}{ll}\dots & a_{1n\hspace{0.17em}}\end{array}\end{array}\\ a_{21\hspace{0.17em}}& 1& \begin{array}{ll}{a}_{23\hspace{0.17em}}& \begin{array}{ll}\dots & a_{2n\hspace{0.17em}}\end{array}\end{array}\end{array}\\ \begin{array}{lll}{a}_{31\hspace{0.17em}}& a_{32\hspace{0.17em}}& \begin{array}{ll}1& \begin{array}{ll}\dots & a_{3n}\end{array}\end{array}\end{array}\\ \begin{array}{lll}\dots & \dots & \begin{array}{ll}\dots & \begin{array}{ll}\dots & \dots \end{array}\end{array}\end{array}\\ \begin{array}{lll}{a}_{n1}& a_{n1}& \begin{array}{ll}{a}_{n3}& \begin{array}{ll}\dots & 1\end{array}\end{array}\end{array}\end{array}\right]$

(3)

$A_{norm}=\left[\begin{array}{lll}\frac{a_{11}}{{\displaystyle \sum }_{j=1}^n{a}_{j1}}& \cdots & \frac{a_{1n}}{{\displaystyle \sum }_{j=1}^n{a}_{jn}}\\ ⋮& \ddots & ⋮\\ \frac{a_{n1}}{{\displaystyle \sum }_{j=1}^n{a}_{j1}}& \cdots & \frac{a_{nn}}{{\displaystyle \sum }_{j=1}^n{a}_{jn}}\end{array}\right]$

(4)

$W=\left(W_1=\frac{{\displaystyle \sum }_{j=1}^n{a}_{1j}}{n},W_2=\frac{{\displaystyle \sum }_{j=1}^n{a}_{2j}}{n},W_3=\frac{{\displaystyle \sum }_{j=1}^n{a}_{3j}}{n},\dots ,W_n=\frac{{\displaystyle \sum }_{j=1}^n{a}_{nj}}{n}\right)$

(5)

$CI=\frac{\left(MaxEigenvalue-n\right)}{\left(n-1\right)}$

(6)

$CR=\raisebox{1ex}{$CI$}\!\left/ \!\raisebox{-1ex}{$RI$}\right.$

(7)

Table 6.

Random consistency index by Saaty. ⁹⁸

N	1	2	3	4	5	6	7	8	9	10	11	12	13
RI	0.00	0.00	0.58	0.901	1.12	1.24	1.32	1.41	1.45	1.49	1.51	1.48	1.56

The linear models for diagnosing each of the eleven febrile diseases using aggregate diagnostic factor index (ADFI) are presented in Table A.3. Table 7 shows the formulated probability ranges, $P_i$ based on equation (8) and generated ADFI, to determine the probability of diagnosing a particular disease, $i$ for a patient for appropriate treatment decision. ¹⁶

$P_i=\frac{ADFI}{5}*100$

(8)

Table 7.

Occurrence probability ranges. ¹⁶

Uniform rating	ADFI range	Probability range (%)
1	0.000 - 1.000	0.01-20.00
2	1.001 – 2.000	20.01 – 40.00
3	2.001 – 3.000	40.01-60.00
4	3.001 – 4.000	60.01 – 80.00
5	4.001 – 5.000	80.01 – 100.00

Discussion of results

Priority weight of symptoms in the AHP model

The normalized mean opinion score of physicians’ responses for symptoms of MAL are shown in Table A.4a. It is noted that the least z score obtained from the physicians’ confidence level for each symptom is 7.09677. Data from Table A.4a was used to construct the symptom-by-symptom pair-wise comparison matrix to determine the relative importance of the symptoms of MAL and the process was repeated for other diseases. Table A.4b shows the symptom-by-symptom pair-wise comparison matrix for MAL with diagonals equal to 1 based on equation (3) while Table A.4c shows the normalized comparison matrix based on equation (4). The generated weighted eigenvector in Table A.4d based on equation (5) shows the relative weight of each symptom for MAL diagnosis, which is the row average of the normalized comparison matrix. In summary, Table 8 shows the overall AHP model indicating the priority weights of the relevant symptoms in the diagnosis of the individual diseases.

Table 8.

AHP model showing priority weights of relevant symptoms in diagnosing each disease.

Symptom	MAL	ENFVR	HVAD	UPUTI	LWUTI	URTI	LRTI	TB	LASFVR	DENFVR	YELFVR
ADPN		0.124443487									0.092987518
BLDFASG									0.139937233	0.099257141	0.083171298
BLDYURN				0.130360236	0.133667084
CGH						0.149591655	0.137124913	0.099346593	0.085700491
CHLNRIG	0.130677553			0.136319003						0.100617305	0.087593301
CHSPN							0.127897961	0.091821641	0.090689627	0.07501831
CLDYURN				0.152927098	0.162152691
DIFBRT							0.138326743	0.086077404	0.09378772
FTG	0.122921985		0.08373634	0.127625064			0.117120261	0.090299418	0.111370403	0.105186064	0.089098664
FVR	0.149200559	0.128716071	0.083308841	0.158981854	0.157596996	0.124565376	0.132007444	0.090672794	0.139736059	0.113068043	0.10556357
GENBDYPN	0.129623398	0.11069223
HDACH	0.136475404	0.122540572								0.104697799	0.092924795
HGDFVR			0.076575734			0.116075038	0.128363185		0.133097288	0.11090573	0.102333312
JNTNMSCPN			0.070056376							0.110661598	0.09157624
LWGDFVR			0.084217276			0.122301286	0.105179499	0.094463991	0.091132212
NGTSWT			0.070831218					0.107532132
NUSNVMT				0.129774128				0.061851081	0.114548966	0.092770202	0.083610362
SRTRT			0.079007134			0.135562384				0.087817808	0.070407075
WGTLS			0.104710503					0.115200689
WHZ							0.113979995	0.062497307
JONPN	0.122021193
BITAIM	0.111118804
PRTN	0.097961104
SWRFVR		0.123797214
DRH		0.106814591
CNST		0.078773517
LTG		0.108143042
GAD		0.096079276
GENRSH			0.095252091
LMPNDSWL			0.090763353
MUTUCR			0.085045555
PNFLBSTR			0.076495578
PNILN				0.164012616
SPPBPN					0.17787234
URNFQC					0.186132666
PNFLURNTN					0.182578223
CTRH						0.150238538
FOLBRT						0.102247918
EPNOTG						0.099417805
LMP								0.10023695
YELOEY											0.100733864

Out of the eleven diseases, only MAL and UPUTI had negative CR of −0.015245828 and −0.003411893 respectively, instead of zero (0). The rest of the diseases had zero (0) consistency ratio values. According to Teknomo, ⁹⁹ negative CR is theoretically not possible but may arise due to error in the approximation of Eigen value especially with Microsoft Excel spreadsheet since Excel gives approximated Eigen values. To validate the weights of the symptoms and resolve these issues, the same sets of values (physicians’ response) were fed into Super Decision Software, ¹⁰⁰ and approximately equivalent values (normalized matrix and weighted eigenvector) as that from the Excel spreadsheet but with CR of zero (0) for all the diseases.

Figure 3 shows a typical spreadsheet result dashboard of ADFI and $P_i$ values obtained for MAL based on inputs from an FHW on the observed symptoms and considered risk factors presented by a patient. Based on medical experts’ advice, a threshold of occurrence $P_i$ = 37was acceptable for diagnosing or predicting the presence of MAL disease.

Figure 3.

Typical result dashboard for disease diagnosis (e.g. MAL).

Performance evaluation of AHP-based MDSS

The developed linear models describing the ADFI for each febrile disease were used in testing 3249 patient data, collected from public and private medical institutions. The patient data comprises patient and physician demographic information, patients’ symptoms, risk factors that predispose a patient to a particular disease, suspected disease by the physician, and the confirmed disease after carrying out laboratory tests. Data used was for patients aged between 5 years and above, which comprises 1336 males and 1913 females.

Further analysis using accuracy, precision, sensitivity, and F1 score are presented in Table A.5. Results indicate that the performance of the developed AHP model⁺ for the MDSS very closely mimics the physician’s diagnosis based on these metrics. In terms of accuracy, the AHP model⁺ achieves these results: MAL (90.06%), ENFVR (85.4%), HIV/AIDs (92.1%), UPUTI (87.5%), LWUTI (85.4%), URTI (86.9%), LRTI (96.9%), and TB (91.4%). Furthermore, the AHP model performs slightly better than the physicians’ decision in the diagnosis of LASFVR (0.685714; 0.470588), DENFVR (0.723404; 0.592593), and YELFVR (0.733333; 0.375), in terms of F1score. Figures 4 and 5 present the accuracy and F1 score plots comparing the physicians’ and the AHP models’ results for each febrile disease. Similar to Maranate et al. ⁸³ who observed nightmares as the most ranked symptom, our study identifies fever to be associated with every considered febrile disease. Samuel et al. ⁸⁶ used Fuzzy AHP to predict heart failure obtaining an average accuracy of 91.10% while our study achieves highest accuracy of 96.9% in the prediction of LRTI. Furthermore, compared to Uzoka et al. ⁵⁸ where AHP method was used to predict typhoid obtaining an accuracy of 78.91%, our study achieves a prediction accuracy of 85.4% for the same febrile disease. The AHP engine for tuberculosis diagnosis in Fergus and Stephen ¹⁴ obtained an accuracy of 82% while the present study achieves an accuracy of 94.1% indicating significant improvements in providing sustainable solution as an information system for febrile disease diagnosis and treatment.

Figure 4.

Accuracy: Physician vs. Model and Physician vs. Model+ diagnosis.

Figure 5.

F1 score: Physician vs. Model and Physician vs. Model+ diagnosis.

Implication of results

The AHP-based MDSS studied here aids both experts and non-experts like physicians and field health workers (FHWs) in diagnostic decision-making, particularly beneficial in rural regions and low- to middle-income countries (LMICs) where medical access is limited. This research extends MCDA’s application, highlighting its ability to mimic human intelligence in decision-making, as evidenced by performance metrics. Additionally, it reinforces previous findings on the influence of risk factors on disease predisposition and transmission. ³⁵ The model exhibits high precision in diagnosing several common tropical diseases such as MAL, ENFVR, HIV/AIDS, URTI, LRTI, UPUTI, and LWUTI, reflecting physicians’ accurate diagnoses, validated by the models. These diseases, frequently encountered in the tropics, result in ease of diagnosis and clinical accuracy. However, for less common diseases like Dengue, Yellow fever, and Lassa fever, which have higher fatality rates and lower disease burden, the model shows lower precision due to reduced suspicion and diagnostic accuracy.

Recommendations

In resource-limited environments, the immediate availability of new or improved healthcare tools is often unfeasible due to significant time and resource investments. Hence, utilizing existing evidence-based solutions becomes crucial for enhancing patient care. Adopting the AHP-based MDSS, along with suitable algorithms or linear models, can significantly improve decision-making and resource allocation. As national standing orders may not adequately address multi-symptom, multi-disease conditions, deploying the results through an app in primary health centers can prevent overuse of RDT kits, reducing overhead expenses and saving government funds. However, in implementing the app, we recommend that the FHWs be equipped with RDT kits to enable more intelligence gathering for further improvement of the services provided by the app. This will also enable faster and more accurate disease diagnosis, lowering morbidity and averting mortalities. Other important recommendations include:

(i) a usability evaluation of the developed system to determine the level of its usefulness, acceptability and user satisfaction;

(ii) a more precise age-specific algorithm should be developed since the causes of febrile illness vary across diverse age groups.

Conclusions

This study addressed the complex task of selecting a suitable weighting method for MCDA, presenting an AHP-based MDSS for diagnosing eleven febrile diseases, considering associated risk factors. Utilizing datasets containing physicians’ experiential knowledge and patients’ confirmed laboratory tests, the model outperformed existing literature in accuracy predictions for certain diseases. Notably, the model closely mirrored physicians’ suspected diagnoses, especially when risk factors were considered. The AHP model achieved high accuracy values (85.4% - 96.9%) for various diseases, suggesting its utility, particularly in resource-constrained settings. Importantly, the model accounted for co-morbidities, enhancing its practicality in real-life situations. Proposing a mobile app with a user-friendly interface, integrating privacy and security features, would facilitate timely and accurate decision-making in diagnosing and treating febrile diseases, potentially reducing the overuse of RDT kits by FHWs in remote communities.