In a first, scientists develop an AI tool to help doctors distinguish between infectious diseases
- The study investigates both statistical and machine learning approaches.
- WHO has categorized dengue as a "neglected tropical disease."
- A prediction tool based on multi-nominal regression analysis and a machine learning algorithm was developed.
Accurate diagnosis is essential for the proper treatment and ensuring the well-being of patients. However, some diseases present with similar clinical symptoms and laboratory results, making diagnosing them more challenging.
In recent years, artificial intelligence (AI) has been used to diagnose numerous diseases such as Alzheimer's, cancer, diabetes, chronic heart disease, tuberculosis, stroke and cerebrovascular, hypertension, and skin and liver disease.
Google Health, for instance, has expanded its research and applications to provide clinicians with "easy-to-use tools" to help them better care for patients. Its Automated Retinal Disease Assessment (ARDA) tool uses AI to help healthcare workers detect diabetic retinopathy.
Google Health has also used an AI algorithm that examines scans of the back of the eye to help predict the risk of patients suffering a major cardiac event.
Now, a group of researchers from the Manipal College of Pharmaceutical Sciences in India has developed a new machine learning-based tool that could help doctors distinguish between different tropical diseases, including dengue and malaria.
The study is the first of its kind wherein both statistical and machine learning approaches were explored simultaneously for differentiating tropical infections.
Detecting tropical diseases has always been challenging for physicians in emergency settings. Four of them, namely dengue, malaria, leptospirosis, and scrub typhus, have clinical symptoms that resemble each other.
"Even an elaborate diagnosis could take three days for the results. That is what prompted us to explore diagnosis with an AI tool," Girish Thunga, Assistant Professor Senior Scale in the Department of Pharmacy Practice at Manipal College of Pharmaceutical Sciences, India, told Interesting Engineering (IE) in an interview.
A tool to distinguish between dengue, malaria, leptospirosis, and scrub typhus
Tropical infectious diseases such as dengue, malaria, leptospirosis, leishmaniasis, scrub typhus, and rickettsial fever are major causes of acute febrile illness, which is a vital concern for the health of people in countries where they are common.
The World Health Organization has categorized dengue as a "neglected tropical disease" (a disease related to poverty and lacking funding for research and development on treatment and cures), despite almost half the world's population living in areas where they are at risk of contracting the disease.
Leptospirosis, malaria, scrub typhus, and endemic fever are among the illnesses that are frequently confused with dengue fever, according to a study of clinical and laboratory profiles of dengue-like illnesses in a tertiary care hospital in West Bengal, India, a region where dengue is endemic.
One main reason for the difficulty in diagnosing tropical infections is that they present with "similar laboratory values, overlapping symptomatology, early asymptomatic presentation, misunderstanding, and delayed diagnosis," noted the PLOS paper.
Non-specific clinical presentations of these infectious diseases also make it challenging to predict wider outbreaks. Moreover, overlapping symptoms and a delay in differential diagnosis could worsen the situation.
The need for a tool that could aid in identifying the early symptoms and distinguishing laboratory parameters of these infections is imperative in reducing their prevalence.
"Early and accurate diagnosis would lead to the right antibiotic [use], thereby reducing antibiotic resistance and [lower] mortality rates in clinical settings," said Thunga.
Our study aimed to make a tool that could distinguish among dengue, malaria, leptospirosis, and scrub typhus in a tertiary care hospital for early prediction."
WEKA software for machine learning(ML) modeling
For a period of nine months, the researchers conducted need analysis at a South Indian tertiary care center.
A nine-item self-administered questionnaire was developed, validated, distributed, and analyzed to estimate physicians' need to differentiate tropical diseases in their setting and components of the same.
While the first part of the questionnaire comprised six questions that were diseased specific - such as the frequency of different tropical infections, the number of cases treated in a week, obstacles in treating tropical infections, challenges in the management of infections, and the perceived need for tool development.
The second part dealt with tool development and included three questions regarding the parameters which clinicians would like to see included, as well as suggested formats for the tool and additional suggestions.
The data for the development of the prediction tool was then collected from the medical record department retrospectively. A prediction tool was then developed, which used multi-nominal regression analysis and a machine learning algorithm.
"A simple scoring system that would differentiate among these infections through a simple decision tree was what we looked at. In total, 800 patients with 200 in each group for the four diseases were collected and analyzed accordingly," explained Thunga.
The Waikato Environment for Knowledge Analysis (WEKA) software was used for machine learning modeling. It was applied to test binary (one disease at a time) and multi-class (all four diseases) classification.
WEKA is an open-source machine learning software in JAVA. It contains tools for data preparation, classification, regression, clustering, association rules mining, and visualization.
The questionnaire that was circulated among 40 physicians and postgraduate students in the department of medicine came as a boon. The doctors claimed that they treated an average of 24 tropical infection cases weekly. They felt the diagnosis was challenging, and the management of symptomatology was difficult.
As per the study, 35 physicians felt the need for the development of a decision-making tool. Thirty-four of them agreed on including laboratory parameters and [35 of them] clinical presentations as the main criterion in the tool.
The questionnaire revealed that dengue, malaria, leptospirosis, and scrub typhus were the most common tropical infections in that setting; with sodium, total bilirubin, albumin, lymphocytes, and platelets the more common laboratory parameters; and abdominal pain, arthralgia, myalgia, and urine output were the clinical presentation were the best predictors of disease.
The tool offered a predictability of 60.7 percent, 62.5 percent, and 66 percent for dengue, malaria, and leptospirosis, respectively, and a 38 percent predictability for scrub typhus.
"The multi-classification machine learning model observed to have an overall predictability of 55–60 percent, whereas a binary classification machine learning algorithms showed an average of 79–84 percent for one vs other and 69–88 percent for one vs one disease category," the research documented.
Challenges and limitations
Thunga stated that the team faced many challenges. "Several permutations and combinations had to be done initially. Every patient would have varying degrees of the range of the disease, and we would come across false positives. Of course, eventually, we worked on it," he said.
He said that further studies were required to provide a detailed insight into the application of this study.
"The credibility of our findings could be further analyzed and upgraded based on the respective clinical scenario," he added.
The paper states that future studies could focus on broad aspects of the disease rather than parameters considered in the study, based on a better knowledge of geographical distribution.
The researchers also warn of the limitations of their study. Some of them include the retrospective collection of data collection which meant that clinical parameters were not recorded during initial visits to the emergency department or clinic, when they may have been more accurate.
In addition, "The findings from this single-centered data cannot be generalized to another part of the world as the nature and presentation of tropical diseases varies from location to location. And as we used the WEKA software for machine learning, it doesn’t provide true negatives, false negatives, and specificity," Thunga noted.