The LalelaLung Study: Digital Stethoscope Clinical Evaluation

Background and Rationale The World Health Organization Integrated Management of Childhood Illness (IMCI) algorithm classifies pneumonia in children with cough and/or difficult breathing primarily on the basis of elevated respiratory rate and chest indrawing. Lung auscultation was historically excluded from IMCI because of its poor reproducibility among non-physician health workers. Since IMCI's introduction, the rollout of Haemophilus influenzae type b and pneumococcal conjugate vaccines has shifted the etiology of childhood lower respiratory infection toward viral pathogens, and placebo-controlled trials indicate that most IMCI-defined non-severe pneumonia is self-limiting. Reliance on respiratory rate alone yields low specificity, driving substantial antibiotic overuse and antimicrobial resistance. AI-enabled digital stethoscopes can reintroduce standardized, objective auscultation by automatically classifying crackles and wheezes with accuracy comparable to expert physicians. The StethoMe device, a CE-marked (EU Class IIa) system using a deep convolutional recurrent neural network trained on more than 25,000 labeled lung-sound recordings, has demonstrated 85-90% agreement with physician reference panels in prior validation and pilot work conducted by the study consortium across multiple low- and middle-income settings.

Overall Study Design The LaLeLa Lung Study comprises two objectives conducted at two primary healthcare facilities in Cape Town, South Africa. Objective 1 is a randomized, triple-blinded, individually allocated diagnostic accuracy study (with nested device-validation sub-studies) evaluating whether IMCI enhanced by an AI-enabled digital stethoscope improves pneumonia diagnostic accuracy and antibiotic targeting relative to standard IMCI. Objective 2 is a mixed-methods, concurrent-triangulation implementation study evaluating usability, acceptability, and fidelity of the digital devices in routine care. The study will enroll a total of approximately 380 participants (350 in Objective 1; up to 30 health workers and caregivers in Objective 2).

Objective 1: Diagnostic Accuracy Study

Objective 1 enrolls 350 children at Site B Clinic, Khayelitsha, randomized 1:1 to IMCI enhanced by the StethoMe AI-enabled digital stethoscope or to standard IMCI care. A computer-generated randomization sequence prepared in advance by the study statistician and implemented through REDCap is used, with stratification by age group (\<1 year and \>=1 year) and allocation concealment from enrollment staff. The design is triple-blinded. Caregivers/participants, routine health workers performing IMCI assessments, and study clinicians performing the digital recordings are all blinded to the device's real-time AI classifications, which are permanently disabled on the device interface for field users. The independent physician reference panel is blinded to study arm, AI outputs, and participant identifiers. Only the statistician holds the allocation key. Importantly, AI outputs do not inform clinical care in either arm, and all participants receive identical study procedures and full IMCI-standard care.

After informed consent and screening, each child is first assessed by a routine clinic health worker who documents IMCI findings and management (including antibiotic prescription or referral) on a study case-management form, without access to the digital stethoscope or study-arm allocation. The child then undergoes an independent structured IMCI-based respiratory assessment by a study clinician, who obtains StethoMe lung-sound recordings at four standardized chest positions. The embedded algorithm computes respiratory rate and classifies abnormal sounds in real time, but all outputs remain concealed. Pulse oximetry (Masimo Rad-G or equivalent), lung ultrasound (Butterfly iQ+), and chest radiography are also obtained. Imaging may be shared with the health worker on request but only after the initial treatment decision and is stored in the regional system using study identifiers. Each enrolled child completes a single in-person encounter (anticipated 60 minutes, integrated into routine clinic flow) followed by a telephone outcome assessment at day 7.

Reference diagnoses are established retrospectively by an independent three-physician panel reviewing compiled, de-identified case records (health worker and study-clinician findings, SpO2, imaging, tuberculosis investigations where applicable, treatment, and follow-up status), excluding any AI output. The panel adjudicates in stages. Stage 1 uses clinical information excluding lung sounds and imaging. Stage 2 adds lung sounds. Stage 3 adds imaging. Blinding integrity is maintained through separation of enrollment, assessment, and follow-up personnel and a weekly blinding-compliance checklist verified by the principal investigator. Unblinding occurs only when essential for clinical management and must be authorized by the principal investigator and documented.

Objective 1: Nested Sub-Studies

Two cross-sectional device-validation sub-studies are nested within Objective 1. In the first, a subset of approximately 225 participants has one chest-position recording obtained in parallel with the AI Diagnostics digital stethoscope (a SAHPRA-approved device for tuberculosis detection). Among children with features suggestive of pulmonary tuberculosis, device classifications are recorded but concealed and not used clinically, with additional 28-day telephone follow-up to support a composite microbiological, radiological, and clinical reference standard.

In the second, the first 100 participants enrolled with the study clinician present undergo additional respiratory assessments with the Perin Health Patch multimodal wearable and the ChARM automated respiratory-rate monitor, with paired clinician respiratory-rate counts and conventional auscultation obtained during sequential timed recordings. Study staff remain blinded to all device-generated outputs.

Statistical Considerations and Sample Size Analyses follow a pre-specified Statistical Analysis Plan finalized before unblinding, conducted primarily on a complete-case/per-protocol basis among randomized participants with an available reference diagnosis, with intention-to-treat sensitivity analyses. Diagnostic performance is summarized using sensitivity, specificity, overall accuracy, diagnostic odds ratio, and receiver operating characteristic (ROC) area under the curve, with between-arm comparisons by two-sample tests of proportions and DeLong's test, and adjusted comparisons by multivariable logistic regression (adjusting for age, sex, baseline SpO2, and symptom duration). The primary sample size of 350 (175 per arm) provides 85% power at two-sided alpha = 0.05 to detect a difference in diagnostic accuracy from 0.67 (standard IMCI) to 0.80 (AI-enhanced IMCI), inflated for an anticipated 10% rate of missing or indeterminate reference diagnoses. The nested sub-studies are powered separately for non-inferiority of recording quality (10% margin) and for respiratory-rate agreement (equivalence margin of +/-3 breaths per minute). Enrollment is expected to require approximately 7-13 months depending on seasonal respiratory illness presentation.

Objective 2: Implementation Study Objective 2 is conducted at Delft South Clinic, Delft, using a mixed-methods, concurrent-triangulation design over a four- to six-week controlled-implementation period. Participating health workers use the StethoMe device with its AI interface visible, alongside the Perin Health Patch, the ChARM device, and the Phefumla 2.0 smartphone-connected pulse oximeter. Device outputs are visible but do not drive clinical decision-making. Quantitative data collection comprises structured observation of device-use fidelity (correct chest positions, workflow adherence, time per recording), workflow integration, technical performance (proportion of successful recordings), and post-encounter standardized usability and acceptability surveys (System Usability Scale). The qualitative component comprises semi-structured and in-depth interviews with approximately 5-10 health workers and 5-10 caregivers, purposively sampled and conducted in the participant's preferred language (isiXhosa, English, or Afrikaans), audio-recorded, transcribed, and translated for thematic analysis. Coding follows a hybrid inductive-deductive approach informed by the Consolidated Framework for Implementation Research and the Technology Acceptance Model, with mixed-methods integration via joint-display analysis.

Risk, Data Management, and Oversight Overall participant risk is no more than minimal because no clinical decision is based on the investigational device outputs, which remain concealed in Objective 1. The digital stethoscopes, wearable patch, respiratory-rate monitor, pulse oximeter, and ultrasound are non-invasive, and chest radiography uses standard low-dose pediatric protocols only when clinically indicated. Electronic data, including de-identified lung-sound recordings, are encrypted and stored on secure, password-protected servers hosted by Stellenbosch University, with clinical and acoustic data captured in REDCap. The study is reviewed and approved by the Johns Hopkins University School of Medicine IRB and the Stellenbosch University Health Research Ethics Committee. All study staff complete a multi-day training program with a competency evaluation before enrollment and quarterly refreshers thereafter.

• Accuracy, sensitivity, specificity, and positive/negative predictive values (Diagnostic accuracy relative to routine health care worker (HCW) diagnosis) [Time frame: Day 1, 7-day follow-up]
• Proportion of correctly indicated antibiotic decisions (Accuracy of antibiotic decision-making) [Time frame: Day 1; 7-day follow-up]
• Accuracy of pneumonia diagnosis (Expanded lung-sound classification accuracy) [Time frame: Day 1; 7-day follow-up]
• Agreement between AI lung-sound classification and physician auscultation [Time frame: Day 1]
• Proportion of digital recordings successfully obtained and interpretable (Feasibility of digital auscultation) [Time frame: Day 1]
• Mean time to obtain standardied chest-position recordings (operational metrics of digital auscultation) [Time frame: Day 1]
• Caregiver-reported chld status [Time frame: Day 7]
• Proportion of routine HCW encounters (HCW fidelity to the IMCI algorithm) [Time frame: Day 1]
• Proportion of interpretable lung-sound recordings (Recording quality - AI Diagnostics digital stethoscope vs. StethoMe (non-inferiority) [Time frame: Day 1]
• Sensitivity, specificity, positive and negative predictive values (Diagnostic accuracy of AI-enabled digital stethoscope for pulmonary tuberculosis) [Time frame: Index visit (Day 1); follow-up up to Day 28]
• Agreement of automated respiratory-rate measurement and breath counts [Time frame: Day 1]
• Proportion of Perin Health Patch lung-sound recordings meeting pre-defined acousic quality criteria (Interpretability of Perin Health Patch lung-sound recordings) [Time frame: Day 1]