Artificial intelligence (AI) is a set of bioinspired algorithms that are used to solve problems in different applications. Within this wide area, machine learning (ML) is a common subfield in which models learn from examples of data, taking advantage of the idea of adjusting parameters in classification or regression tasks (1). There are several different ML models according to the fundamental concepts for adapting the parameters, with diverse examples including naive Bayes, decision or classification trees, support vector machines (SVM), and artificial neural networks (ANNs), which emulate the behavior of the brain through connectionist models. Besides these and other ML models, new models are continuously being proposed (2).

Tuberculosis (TB) is a disease caused by the Mycobacterium tuberculosis bacillus, and the World Health Organization still considers it a global emergency because of its high estimate of more than 1.4 million fatalities in the last 3 years (3). In developing countries, TB incidence is as high as 282,000 new cases in recent years with a mortality rate of 2.4 per 100,000 populations. In one specific place, Colombia, the reported TB incidence was 33, the prevalence was 48, and the mortality was 1.6 per 100,000 populations. Given these numbers, any contribution to decreasing TB fatalities is welcomed. M. tuberculosis is slow-growing and replicates itself every 24 h, an important fact that determines subacute symptoms. Additionally, the main organ affected by TB is the lung, and because of this, the main signs of the disease are respiratory-related (3). Coughing and expectoration allow for assessing the probability of TB by studying sputum; however, because TB is an infectious disease, the accurate diagnosis is microbiological (4).

In the health area, AI has been applied to solve problems in public health, medical images analysis, and diagnosis support systems (5–8). For TB, different approaches have been proposed since 1999 with the work of El-Solh et al. (9), for whom medical images were the main source of information. Advances in this field have allowed for better detecting thoracic diseases including TB, pneumonia, asthma, and cancer (10, 11). Investigators have widely used specific ML models in health systems to contribute to improving TB diagnosis by taking advantage of available meaningful data (12, 13), such as data from clinical information (14–16), or molecular biology (17, 18).

ANNs have been particularly valuable in incorporating ML into TB diagnosis through different architectures such as multilayer perceptrons (MLP), self-organizing maps, and adaptive resonance theory (ART) joined to fuzzy models in the Fuzzy-ART approach to support detection and clustering in risk groups for pulmonary TB (19–21) and pleural TB (22–24). Researchers have used different data sources to support health professionals in daily tasks such as collecting breathing acoustic signals (25) and other clinical variables (20, 26).

Finally, TB researchers have used deep learning (DL) architecture using vast data sets to provide scenarios based on images (27–29). For instance, one important task was establishing the ImageCLEF data set, which allowed users to determine TB type and treatment resistance using coaxial tomography images (28, 30); researchers have also used images from radiography to support health professionals' decision making (31–33). Generally, DL has been widely applied in assisting with medical diagnosis, utilizing radiography images, and obtaining highlight results (34, 35). Additionally, one DL subfield, transfer learning, entails refining large pretrained models with new data, and several researchers have applied transfer learning to the same kinds of medical images (27, 36).

Nevertheless, despite its demonstrable benefits, ML's effectiveness can be limited by data availability constraints related to inadequate information technology infrastructure. Precarious health systems that cannot or do not collect radiographic information or conduct specialized testing significantly complicate the implementation of ML models. Researchers have analyzed these characteristics and proposed infrastructure for developing regions that can accommodate few variables and poor information systems have been treated for developing regions (19, 21).

The present work proposes ML techniques as a tool in the loop of TB diagnosis, where health professionals make decisions but with extra help based on limited available data. This scenario is studied for using ML in situations with limited infrastructure for application within the complete TB diagnosis protocol.

The concept of the “algorithm-in-the-loop” is related to the use of ML models to support decision making and improve both human–computer interactions and human performance (37). Interaction between the model and users in a loop is not limited to simple representations of performance such as numbers but extends to a global idea that articulates ethics, policies, and standards (38). Including AI and ML stages in the clinical decision making support workflow can ultimately improve patient experiences and outcomes and optimize health system performance (8). Interactive ML is another term for when algorithms and humans work together to improve the results in terms of metrics, understandability, and outcomes (39).

For the case of TB, diagnosis was long based on respiratory symptoms followed by testing suspicious patients with a serial sputum smear; however, although this test is simple, it is necessary to consider some aspects in determining its usefulness. Smear microscopy is performed using sputum smear and staining that allows direct microscopic visualization of the bacillus. However, diagnostic sensitivity is low, around 60%, because a high number of microorganisms per cubic millimeter of a sample is required to obtain results (40). Indeed, a high percentage of people with the disease cannot be diagnosed using this method, and furthermore, detected bacillus could be a non-TB mycobacterium. A more sensitive assay is a culture in either solid or liquid medium, which needs at least 2 weeks to obtain results (41). Following more recent advances, molecular testing is now available: Polymerase chain reaction (PCR) identifies the TB bacillus with high sensitivity and in approximately 2 h (42). However, the infrastructure for this technology is limited in developing countries such as Colombia.

From the ML point of view, different applications have particular characteristics such as requiring biomedical data that have high uncertainty and incompleteness (43), and strategies beyond straightforward ML are sometimes demanded. For the present study, ML in the loop (MLL) is investigated; this strategy depends on how the ML tool will be used. Researchers have analyzed the necessary workflows to improve results (44), but in medicine, where health professionals play an indispensable role, other investigators have studied the doctor-in-the-loop in terms of system performance (45, 46). Today, how ML models perform is no longer the sole concern; models' generalizability and functionality during human interaction are also important. Assessing these broader aspects of performance allows for understanding important aspects of decision making and operation that must be considered in system designs (47).

Figure 1 depicts the MLL process for TB diagnosis support that was studied for the present work. First, a subject with respiratory disease symptoms arrives at the medical center for either a consultation or an emergency. There, a member of the medical staff examines the possible patient and then sends the patient to internal medicine for a more detailed examination. After this deeper analysis, if the patient's respiratory symptoms continue, medical staff request three main exams to detect pulmonary TB: sputum smear microscopy, sputum culture, and molecular assay (GenXpert®). If results from these three exams indicate infection, the patient begins antituberculosis therapy. Meanwhile the results are definitive, there is no positive diagnosis. However, the patient initialize the antituberculosis treatment. It is at this point where ML was applied to assist the medical staff members in diagnosis.

At the study facility, the health care workers are responsible for acquiring basic patient information equivalent to the medical records obtained in other stages. This information is input into a registry for the use of the institution's TB program; the protocol to detect TB can be time-consuming, and using ML with this registry could expedite diagnosis. This study proposed to apply MLL searches to support health care workers during the time the test results take. This allows staff to efficiently manage patient treatment according to the need for isolation, hospital capacity, and necessary medications.

Table 3 shows the findings for the training process and the test scores with data from the year left out in the cross-validation described before; accuracy (ACC), sensitivity (SE), and specificity (SP) were collected to determine the differences due to the balance between positive and negative TB for each year (see Table 2). Additionally, the area under the receiver operating curve (AUC) allowed for considering SE and SP simultaneously.

The LR, RF, and MLP models achieved the best results, obtaining the highest AUC, 0.84, in the test set (see Table 3). This value can be compared with the maximum AUC of 0.96 in the DT model for the training set, demonstrating that it was difficult to generalize the findings from the present application.

Table 4 presents the ACC, SE, SP, and AUC means and standard deviations for the three test data subsets. The table shows that MLP obtained the best results for ACC, SE, and AUC and that SP was the best with the LR model. These findings suggest that combining models might give better results for these metrics. Nevertheless, although SP was the best with the LR, that model had the worst results for ACC and SE, which suggests this model's suitability for the objective task of finding negative TB cases. Finally, the SVM model gave the worst results for most metrics.

Table 5 presents the best results for each metric for all the studied models and the full data set, showing that the LR model had the best accuracy, SVM had the best sensitivity, and MLP had the best specificity. Additionally, following subsection 3.3, all models were checked for relevance. Specifically, for each model, the input variables (see Table 1) were set at 0, and then, ACC, SE, and SP were computed. Figure 2 shows the effect of this processing, notably that type of population was not important in the LR, RF, and MLP models; when the zero values were eliminated, the models' performance improved. Figure 2D shows that age caused significant differences in the SVM model. Finally, all variables were relevant in the MLP model.

Table 6 presents the findings from testing aML and TPOT, which require less intensive user exploration of the hyperparameters. The table shows that the automated ML was more successful than manual exploration (see Table 3), although the results were similar. The first model, for the year 2019, applied six ML models: two passive-aggressive, two MLPs, one extra tree, and one gradient boosting. The second model, for 2018, had 28 models that included a number of the different strategies presented here (e.g., MLP, RF, and logistic regressors). For the 2017 case, aML produced a combination of five models (two random forests, one mlp, one passive-aggressive, and one stochastic gradient descent). Table 7 presents the aML and TPOT results for all 3 years. Specificity is considerably affected in this automatic generation of models, which is ineffective and not appropriate in the context of diagnosis support.

TB detection in earlier stages is important to prevent transmission of the disease. However, irrespective of when a patient is diagnosed, patients in the populations studied in this work must be kept in isolation because these patients tend not to maintain safe distances as they are being treated.

Because of the lack of specific clinical symptoms, it is difficult for physicians to diagnose tuberculosis, but meanwhile, patients require rapid isolation to prevent spreading the disease to others. Presumptive TB cases require further analysis, and tools for completing specific tasks could reduce the workloads of health professionals. ML and AI could be effective in this context while keeping decisions under the purview of the medical staff. Furthermore, in developing or low-income countries such as Colombia, ML and AI can extend the availability of health care to remote regions with limited infrastructure and few if any health care personnel.

There remain many challenges to applying ML and AI in the health informatics field, but doing so can contribute to easing burdens for clinical personnel; further testing of these applications in real-world settings will be highly beneficial. Furthermore, the coworking between health professionals and health care AI is a challenge. The American Medical Association calls for considering AI an augmentation to human intelligence rather than a replacement (62). Recent authors have reported on developing this kind of articulation with health professionals as the center of the entire strategy (12).

In this study, the high incidence rate in the analyzed data set was related to the stage of the diagnosis process, although despite this, it is possible to see that not all presumptive TB cases were ultimately diagnosed as positive TB. This indicates that the ML tool identified variables that were imperceptible to humans, which could help improve therapy management as well as increase the efficient allocation of clinical resources (time, professional staff, medicaments, space, etc.). However, it was determined in this study that the unbalance between positive and negative TB cases could be offer a difficulty of the ML models training (59). However, the RF, LR, and MLP models achieved similar results for SE and SP, consistent with earlier findings for MLP models (19, 21, 33, 55); these findings support RF, LR, and MLP as appropriate models for diagnosis support. In the present study, MLP had the best AUC metric, which exhibits best balance between SE and SP. Additionally, the proposed models can decrease the number of cases for which treatment begins without a confirmed diagnosis, which should decrease health system costs in time and other resources. Regarding aML and TPOT, finding the hyperparameters was not a dilemma, but the SP results were not as good as they were with other models. Furthermore, it is common for health informatics applications to have access to only small data sets or represent only rare events, and these conditions significantly reduce the accuracy of the results from aML approaches (60, 61).

Diagnostic algorithms have been incorporated into several national and international recommendations and guidelines for optimizing patient approaches. In the case of Colombia, health entities must notify the alert surveillance system of public health diseases, to epidemiologically monitor and clinically control TB to verify the success of the treatment. National TB registries allow for acquiring adequate global information on all the current clinical and sociodemographic aspects of TB as well as the success of the treatment strategies used.

In terms of limitations of the present study, there was a high incidence of TB in the data set, which could have induced bias in the analyzed data; addressing this will require more specific scenarios that involve clinical observation. Additionally, TB culture is considered the gold standard for diagnosis in some cases, especially when the infrastructure of GenExpert is not available. In this study, although the hospital database can only hold a limited number of patients, the HSC is an important center for TB treatment in Bogotá City; future researchers could incorporate data from more institutions that treat TB. Finally, researchers could incorporate more technical aspects such as including ensemble methods, combining different ML models, and considering more sophisticated models as the next steps.

The findings of this study make it possible to conclude that sensitive ML algorithms can support TB diagnosis by considering the clinical features of the cases as well as medical and sociodemographic risk factors of the patients. TB continues to be a global leading cause of death, and challenges remain in identifying, treating, and containing the disease in several communities. The mycobacteria–host relationship can delay diagnosis for a host of reasons, as can limited clinical resources for diagnosis. Computational tools such as those studied here can support timely TB diagnosis and treatment.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

Conceptualization: AO-C and CA. Methodology, supervision, and resources: AO-C and AJ. Software, writing—original draft preparation, funding acquisition, and visualization: AO-C. Validation: AO-C, CA, EV, and AP. Formal analysis, investigation, and writing—review and editing: AO-C, AJ, CA, EV, and AP. Data curation: AO-C, CA, and AJ. Project administration: AJ. All authors have read and agreed to the published version of the manuscript.

This research was funded by Ministerio de Ciencia, Tecnologia e Innovación of Colombia—Minciencias, grant number 123380762899.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.