Artificial intelligence advancement in addressing cough

Abstract

Cough assessment is a vital component of clinical management for pulmonary disorders, yet effective tools to execute quantification are insufficient. The dominant approach incorporates acoustic evaluation, primarily based on cough sounds. Incorporating artificial intelligence (AI), particularly machine learning and deep learning, is poised to revolutionize this field. Although AI’s application is focused mainly on curative cough management, there is modest yet growing interest in using this technology for preventive medicine. However, these advancements are accompanied by technical, ethical, and legal challenges that must be resolved to leverage their full potential. Furthermore, to ensure reliability and address concerns of medical practitioners about technological replacement, the human-in-loop model should be advocated. This overview summarizes recent advancements in AI for cough medicine, covering its current clinical applications, existing limitations, strategies to address these challenges, and future opportunities.

Key Words: Acoustics; Artificial intelligence; Cough; Cough sound analysis; Deep learning; Machine learning

Core Tip: Evaluating cough for pulmonary conditions relies substantially on acoustic measurement, yet existing quantification tools remain suboptimal. Integrating artificial intelligence (AI), particularly machine learning and deep learning, offers a promising pathway for both therapeutic and preventive applications in cough medicine. Current implementations remain confined to weak AI’s predefined roles, while advances in artificial general intelligence hold the potential to overcome these adaptability constraints. Moreover, harnessing current potential requires resolving key technical, ethical, and legal issues by integrating explainable AI, multimodal hybrid approaches, and robust accountability measures within human-centered frameworks.

INTRODUCTION

As a defensive mechanism of the respiratory tract, cough is crucial to maintain airway clearance. It is manifested as a sudden, forceful expulsion of air. Although a normal phenomenon, a chronic or persistent cough may signal a potential health threat mandating further evaluation[1,2]. Cough itself is not an autonomous disease; rather, it commonly mirrors an underlying pulmonary pathology. As respiratory disorders are among the principal drivers of the global burden of disease, the study of cough is of paramount significance[3-5]. Nearly 5%-10% adults are affected by chronic cough with its vexing sequelae such as sleep disturbances, dysphonia, syncope, and urinary incontinence, ultimately impairing their quality of life[6,7]. Chronic cough persists beyond 8 weeks in adults and 4 weeks in children[6]. A wide range of diseases with multisystem involvement may present with cough, as shown in Table 1.

Table 1 Etiology of cough[4-6].

Acute cough (< 3 weeks)

Respiratory tract infections (pharyngitis, tonsillitis, bronchitis, pneumonia)

Subacute cough (3-8 weeks)

Post-infectious cough

Chronic cough (> 8 weeks)

Upper airway cough syndrome

Asthma (cough-variant asthma)

Eosinophilic bronchitis

Gastro-esophageal reflux disease

Drug induced (angiotensin converting enzyme inhibitor)

Chronic obstructive pulmonary disease

Bronchiectasis

Environmental and behavioral factors (smoking, pollution)

Neuronal hypersensitivity

Traditionally, the initial approach to cough begins with a clinical assessment and basic diagnostic investigations, which guide towards empirical therapy followed by sequential trials for non-responders, as depicted in Figure 1[6]. Fueled by the latest technological breakthroughs, cough medicine has evolved from a conventional approach to a modern one that incorporates computational pulmonology. The resource limitations and access challenges highlighted by the coronavirus disease 2019 (COVID-19) pandemic underscored the need for novel diagnostic solutions. Under similar circumstances, innovative technologies such as artificial intelligence (AI) can help overcome these challenges[8,9]. AI serves as a valuable resource in healthcare to strengthen predictive diagnostics, therapeutic efficacy, precision medicine, and predictive prognostics[10,11].

Figure 1 Flowchart showing approach to cough[6]. AAFB: Alcohol and acid-fast bacilli; CT: Computed tomography; FeNO: Exhaled nitric oxide fraction.

AI refers to the ability of a machine to perform a function that mimics human cognition. It can be categorized into three stages based on intelligence level: Artificial narrow intelligence, artificial general intelligence (AGI), and artificial superintelligence (ASI), as illustrated in Figure 2. Basically, narrow or weak AI is designed for a specific domain, which is intellectually inferior to humans. Likewise, general or strong AI is regarded as functionally analogous to human intelligence, and ASI is seen as surpassing human capabilities. As of today, operational AI is limited to a narrow domain, while AGI and ASI still remain hypothetical. The core application of modern AI, known as machine learning, enables machines to learn autonomously from data and perform pattern recognition for instantaneous decision-making without being explicitly programmed. Based on how algorithms learn from data, they are broadly segmented as supervised and unsupervised learning. The niche functions in machine learning involve deep learning (DL), which relies on multiple deep neural networks, allowing it to handle complex patterns and construct sophisticated predictions. Convolutional neural networks (CNNs) and recurrent neural networks (RNNs) are widely used DL models[10,12-14].

Figure 2 An overview of artificial intelligence. The core concept of artificial intelligence is to make machines smart. Based on intelligence levels, the pyramidal hierarchy is maintained, with the artificial super intelligence at the highest rank. The relationships between artificial intelligence, machine learning, and deep learning are also displayed[12-14]. AGI: Artificial general intelligence; ANI: Artificial narrow intelligence; ASI: Artificial super intelligence; DL: Deep learningl ML: Machine learning.

As AI has been revolutionizing various sectors, even healthcare is successfully capitalizing on its rewards. The field of AI officially began in the 1950s, when the term was first coined by John McCarthy[13,14]. During the 1970s, AI was introduced into the healthcare system through the MYCIN program to assist physicians in diagnosing infections and recommending antibiotic therapies[15]. Throughout the foundational era of medical AI and its evolution in the 1980s, INTERNIST-1; its improved version - Quick Medical Reference; and DXplain, were engineered for diagnostic assistance[16-18]. In modern medicine, the transition of AI from a rule-based system to a data-driven system represents a tipping point, particularly since the 2020s, with refinements in DL effectively neutralizing previous challenges[19]. Major advances in healthcare include: IBM Watson, DL-enabled diagnostics, autonomous AI systems, AlphaFold 2, and an automated photoelectrochemical biosensing platform[19-23].

Following these groundbreaking advances in healthcare, AI is driving a massive surge in demand. Its integration into modern healthcare is undeniable because of its potential to handle overwhelming volumes of medical data and mitigate the global shortage of healthcare professionals, which is crucial for expanding personalized medicine. The continuous improvements in AI algorithms and growth in computational power further accelerate its utility. The core assets of the AI system are its capacity to generate economical and accessible solutions through non-invasive, user-friendly modules. It further supports telehealth services, concurrently safeguarding against the spread of infectious diseases. Generally, the applications of AI enhance screening methodologies, diagnostic accuracy, treatment efficacy & effectiveness, prognostic reliability, and operational healthcare management, thereby broadening the availability of comprehensive healthcare[12,19,24-26]. As the modern healthcare landscape inclines toward individualized care, AI’s role in precision or personalized medicine remains crucial. It optimizes clinical outcomes by minimizing adverse effects by analyzing genetic profiles, medical data, and environmental factors. Beyond clinical care, AI plays a vital role in predictive medicine, enabling the anticipation of disease progression and recommending preventive approaches, thereby reinforcing public health preparedness. AI further enhances biomedical innovation by streamlining clinical trials[19,23,27,28]. Advancements have enabled AI penetration across the majority of healthcare domains, where integration with cough medicine is no exception. The COVID-19 pandemic served as a significant catalyst, propelling AI technology into cough medicine[8,29].

Cough is merely a manifestation of an underlying disease, as shown in Table 1. It often coexists with collateral features, including constitutional symptoms, runny nose, sneezing, nasal & throat irritation, chest tightness, dyspnea, and heartburn, as well as adventitious breath sounds. Further evaluation requires determining the underlying condition, planning therapeutic intervention, and predicting the potential outcome[6,30]. Harnessing electronic health records (EHRs), imaging, and genomics to guide AI enables rapid advances in patient care, as shown in Figure 3. Computer vision (CV), natural language processing (NLP), and reinforcement learning are major domains of DL. CV primarily uses CNN architectures to interpret spatial data, such as images and videos, in medical imaging. Likewise, NLP uses RNN modules to process sequential data, such as text, speech, and time-series data in EHRs[13,24,26,31].

Figure 3  Artificial intelligence framework in cough assessment[24,26,31,84].

AI-POWERED DIAGNOSTICS

AI enables objective quantification of cough, eliminating the former reliance on subjective assessment, which was limited to questionnaire-based severity scoring[32]. Acoustic cough analysis is the central paradigm of AI driven diagnostic methods[29]. However, to ensure a robust diagnostic framework, a multimodal approach can be adopted, including acoustic features (cough, voice, speech, breath sounds), imaging, and chemical profiles.

Acoustic biomarkers

Cough sound analysis: By applying signal processing and machine learning, the acoustic properties of cough sounds can be analyzed in detail to yield clinically relevant information. These properties chiefly rely on their temporal (duration, amplitude) and spectral features (frequency/pitch)[31,33,34]. Physiologically, cough is divided into three sequential phases: Inspiratory phase, compressive phase, and expulsive phase. The central element of cough sound analysis is the expulsive phase, which is further divided into explosive, intermediate, and voiced phases, as shown in Figure 4[1,34,35]. Distinctive acoustic signatures are generated by various diseases, including asthma, pneumonia, bronchitis, chronic obstructive pulmonary disease (COPD), COVID-19, and tuberculosis. Additionally, disease severity can be assessed by monitoring cough counts[36-40].

Figure 4 Schematic representation of different phases of cough: Inspiratory, compressive, and expiratory phases. During a cough, the vocal cords also move through these phases, where they are open during the inspiratory and expiratory phases, while they are closed during the compressive phase. Expiratory phase, the audible part of the cough, is further segmented into explosive, intermediate, and voiced phases[34,85].

The outcomes of cough sound analysis may deviate slightly due to the combination of diverse devices (smartphone, sensor, computer) and various ML models (logistic regression, support vector machines, deep neural networks). Furthermore, the scope of cough sound analysis may be tempered by challenges such as: (1) Perceptually identical sounds including non-cough respiratory sound (throat clearing) and non-respiratory environmental noises; (2) Cough triggered by non-pulmonary causes; and (3) Significant disease overlap - obscuring the interpretation. However, these setbacks can be resolved using advanced AI incorporating denoising mechanisms and robust generalizability[41-45].

Standardized acoustic feature extraction from cough excludes the inaudible inspiratory and compressive phases. Nevertheless, these preliminary stages are critical elements in determining cough strength and potential dysfunction. Thus, enabling AI to incorporate these factors to extend diagnostic precision[34,46].

Vocal biomarkers: The application of vocal biomarkers (voice/speech analysis) in pulmonology remains highly specific, focusing mainly on remote monitoring, identifying acute exacerbations, and early intervention for a few conditions (asthma, COPD, COVID-19). However, it holds immense potential beyond its current limitations as AI continues to evolve[47-50].

Breath sound analysis: Acoustic variances in breath sounds (oral, nasal, lung sounds) may serve as a beneficial diagnostic tool. This helps distinguish normal from adventitious breath sounds (stridor, wheeze, crackles)[51,52]. Moreover, it facilitates the identification of subtle breath sounds, primarily observed in children. Currently, breath sound analysis is addressed mainly in pediatrics due to limited cooperation and atypical pulmonary features. Similarly, the studies show overt favoritism towards the analysis of lung sounds. Thus, it is necessary to widen its scope by including adults and incorporating oral and nasal sounds for analysis. Furthermore, to ensure diagnostic accuracy, robust noise-reduction techniques must be implemented to mitigate acoustic contamination[53-57].

Radiomics

Radiography is another area where AI has demonstrated superior performance. It has enhanced medical imaging by improving the detection, characterization, and quantification of abnormalities, particularly in thoracic imaging. It has been boosting efficiency and diagnostic sensitivity. Tools, such as computer-aided detection, also enable scalable screening in resource-limited settings, reducing workload. However, human supervision remains crucial as full automation exceeds current limits. As progress continues, its role in image interpretation is expected to expand[58-62].

Metabolomics

Researchers identified distinct metabolic signatures between healthy, acute COVID-19, and post-coronavirus disease (COVID) patients. This metabolomics approach to COVID could be further expanded with AI to significantly enhance diagnostic value for other respiratory illnesses[63-65].

AI provides a holistic approach to address cough, enabling the identification of non-respiratory etiologies that are often overlooked. Furthermore, it should be emphasized that existing AI tools in cough medicine like cough sound classification, disease detection and automated counting, function within the domain of weak AI. Despite their expertise in targeted medical roles, these applications lack the flexibility to function beyond their targeted roles[39,43,66,67]. Nevertheless, we anticipate that technological advancement toward AGI will overcome existing boundaries, empowering systems to generalize their capabilities.

AI-POWERED THERAPEUTICS

Integrating AI, particularly machine learning and DL, has led to significant improvements in pulmonology. It serves as a core framework in the advancing field of precision medicine, an approach that prioritizes individualized patient care over conventional methods. AI is uniquely capable of integrating and analyzing vast amounts of patient data from multiple sources such as EHRs, medical imaging, and genomics. Moreover, it facilitates prompt intervention through diagnostic accuracy, disease surveillance, and determination of therapeutic effectiveness. Nevertheless, the existing therapeutic capabilities are poised for expansion[10,68-71].

AI-POWERED PROGNOSTICS

Currently, AI aids prognosis in cough medicine by analyzing cough sounds. However, its future scope is expected to integrate additional biomarkers[72].

Beyond expanding the utilization of AI in curative cough medicine, preventive medicine is newly entering the spotlight. AI - driven technologies optimize respiratory disease management by integrating predictive modeling with real - time monitoring to enhance both preventive and corrective actions. It can further interpret epidemiological trends and datasets, thereby, estimating outbreaks and identifying early warning signs. This facilitates strategic preventive measures and educational initiatives to downregulate disease prevalence. Thus, anticipating to achieve deeper penetration into the clinical framework[72-75]. Some major breakthroughs in AI integration in cough medicine are listed in Table 2.

Table 2 Major breakthroughs in artificial intelligence integration.

Target	Data type	AI technology	Findings
Screening tool for TB, COVID-19, pneumonia	Cough sound	AudibleHealth AI (RAISONANCE)[86]	Disease	Sensitivity	Specificity
			Tuberculosis	90%	72%
			COVID-19	97%	75.5%
			Pneumonia	94%	73%
Muti-disease smartphone diagnosis	Cough sound	ResAppDx-EU[87]	Respiratory disease	Sensitivity	Specificity
			Pediatric	83%-97%	82-91%
			Adult	83%-89%	84%-91%
Automated cough analysis	Cough sound	Convolutional recurrent neural network[88]	97.2% accuracy; 92.9% sensitivity; 97.6% specificity
AI cough classifier	Cough sound	Machine learning (mobile based cough sound classifier)[43]	Detection of respiratory diseases (TB, asthma, COPD) in rural Tanzania
Chronic cough identification	Cough with electronic health record data	XG boost model[89]	ROC-AUC of 0.916. Best performance when compared with other machine learning algorithms based on logistic regression and neural network approaches
Lung sound analysis	Lung sound	Feelix smart stethoscope (Sonavi Labs)[90]	Advantages over traditional stethoscope: Quick auscultation since precise positioning is not required; active noise cancellation; automatic abnormality detection
Bio-acoustic foundation model	Cough, speech & breath sound	Google HeAR (health acoustic representations)[91,92]	HeAR outperformed several audio encoder baselines such as TRILL, FRILL, BigSSL-CAP12, and CLAP by achieving highest overall performance across 33 tasks. Ranked as the top model in health acoustic detection while dominating cough inference and spirometry achieving the highest mean reciprocal rank (0.708)
Pattern classification and severity mapping	Cough sound	Swaasa platform[66]	90% accuracy
Monitoring tool	Cough sound	Hyfe cough monitor[93]	90.4% sensitivity (supports from initial screening to patient management)

AI: Artificial intelligence; CLAP: Contrastive language audio pretraining; COPD: Chronic obstructive pulmonary disease; COVID-19: Coronavirus disease 2019; HeAR: Health acoustic representations; TB: Tuberculosis; SSL: Semi-supervised learning.

The power of AI is a double-edged sword because its inherent challenges serve as a vital barrier to its advancement. These challenges primarily involve technical aspects and ethical considerations. The primary concern is data quality, which also requires addressing issues of appropriate patient care, data privacy, and security. Furthermore, the lack of transparency, intrinsic to AI - known as the black box phenomenon, compromises generating diagnostic accuracy, effective treatment, and prognostic reliability. The combination of these issues leads to catastrophic consequences in healthcare delivery, thereby prompting legal inquiries and concerns about responsibility and accountability. Beyond that, clinician’s overreliance on AI promotes cognitive offloading, invariably leading to automation bias and professional deskilling. Additionally, amid rising concerns about job security, clinicians may compromise their professional integrity and diminish the standard of care[76-80].

Evolving beyond AI’s limitations demands the adoption of a comprehensive, human-centric strategy that embraces system optimization and resilient ethical governance. Technical refinement should be applied using a hybrid and multimodal approach, as it integrates multiple internal architectures and diverse datasets (clinical records - text, imaging - images/videos, acoustic biomarkers, and chemical biomarkers) to offer optimal generalization, while ensuring data privacy and security[79-81]. Besides, the black box effect of AI mandates the implementation of explainable AI to provide transparent justification, whose strategies facilitate clinical interpretability by revealing the key features, signals, or patterns influencing model outputs. During cough analysis, these techniques are employed to identify acoustically significant patterns while mapping them to underlying respiratory pathophysiology[82,83]. Furthermore, effective legal and policy frameworks to enforce accountability for its output should be formulated. Ultimately, incorporating a human-centric model is a key strategy that leverages technology to strengthen clinicians’ expertise, thereby protecting them against employment fragility. Despite the significant advancements AI has achieved within cough medicine, extensive future research and clinical trials are imperative to operate at its full capacity[79-81,83].

CONCLUSION

The vast domain of AI covers a spectrum of fields, where cough medicine remains no outlier. Integrating AI remains crucial across the entire clinical framework from initial screening to final prognosis. Though AI poses certain limitations and sparks clinician apprehension about job insecurity, its core value lies in its ability to support clinical judgment and decision-making. With further technological advancement, these limitations are anticipated to be overcome. Thus, AI-human collaboration must be sustained to solidify the transformative potential of AI. Moreover, existing legal obscurities surrounding accountability warrant robust policy prescription. Cough medicine has been yielding promising results so far, though current implications are confined to narrow domain. Nevertheless, future evolution toward AGI holds transformative potential by providing the adaptability required to extend its capabilities beyond predefined roles. Besides, further trials and studies are essential to nurture and strengthen its effectiveness.