Mobile phone auscultation to delineate pneumonia from other respiratory conditions and controls: a prospective cohort study

Abstract

Background

Seeking to develop technology for medical countermeasures in novel infectious pandemics, this funded project intended to use mobile phone auscultation (MPA) to enable frontline detection of respiratory disease for situations with limited diagnostic options. The overall objective of this research was to determine the feasibility of obtaining mobile phone recordings in Emergency Departments with the intention of creating accurate models of pneumonia diagnosis.

Methods

This prospective study enrolled participants in five predetermined cohorts: influenza A, pneumonia, acute bronchitis, other respiratory illness, and controls. Potential subjects were approached in one of three emergency departments (two urban and one rural). Recordings were obtained with mobile phones with no hardware modifications. Recordings sampled bilateral 5th intercostal sites during normal and deep respiration, along with one supraclavicular fossa egophonic recording. Recordings were analyzed with computational nonlinear dynamics which, when applied to auscultation data, represent biofluid dynamics. Maximal Lyapunov Exponent (MLE) and Correlation Dimension (Dcorr) were evaluated to confirm the presence of low dimensional chaos and applicability of Time Series Dynamics (TSD) modeling. Train and test sets were created by 80/20 random sampling of clustered records. TSD models were fitted separately to recordings and then combined into a single “composite” model. Binary classifiers were fitted from extracted features by using logistic regression.

Results

From Nov 21, 2023 to June 12, 2024, 292 subjects were enrolled (64 pneumonia, 59 influenza A, 38 acute bronchitis, 68 other respiratory, and 63 controls) with a mean age of 49 years (SD 17.2). Half of the subjects were male, 68.0% white and 29.2% black. Compared to all others, pneumonia subjects were older (58 years vs. 46 years), more likely male (56% vs. 48%), more likely white (77% vs. 66%), and more likely to have a tobacco history (77% vs. 68%). No completed recordings were excluded from the recording analysis. TSD modeling of egophonic, right normal breathing (RNB), and composite models all produced accurate results. In test analyses, the RNB (sens 85%, spec 81%) and egophonic (sens 85%, spec 86%) models performed best. The composite model achieved sensitivity 91%, specificity 87%, and area under the curve of 89%. The test models yielded one false negative and seven false positives.

Conclusion

This study found that modeling of mobile phone auscultation recordings yielded excellent sensitivity, specificity, and area under the curve to delineate pneumonia from other respiratory illness and controls. Similar models could increase access to accurate diagnosis of multiple medical conditions, particularly for medical counter measures during disease outbreaks. Phase 2 will further characterize models for deployment in larger populations.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12890-026-04135-z.

Introduction

Pneumonia continues to cause significant morbidity and mortality both in the US and globally. According to a systematic analysis, lower respiratory tract infections killed 2.18 million people globally just in 2021. More than 500,000 of these deaths occurred in children younger than 5, and more than half of those deaths took place in countries with low socio-demographic index [1]. From 2001 to 2014, the US saw more than 20 million admissions for pneumonia, with hospital mortality rates around 7% and hospital charges exceeding $80 billion in 2014 alone [2].

Despite the continued morbidity and mortality burden of pneumonia, healthcare continues to face challenges with accessible, accurate, and reproducible diagnostic modalities. While the definitive diagnosis of pneumonia is enhanced by advanced imaging technology, offices, urgent care, and especially low resource counties and countries lack this equipment. Regarding emerging infectious diseases (EIDs), novel pathogens by definition have no diagnostic test available. Countries around the world seek technology that can surveil for new threats, but “most countries do not ave sufficient resources or capabilities to detect, analyze, and respond to emergent biosecurity events that have the potential to spread worldwide.” [3].

Mobile phone auscultation (MPA) could fill these gaps in diagnosis of both community acquired bacterial pneumonia and novel EIDs. MPA would have particular utility in rural and low-income populations, some of which have as many phones as people [4]. Multiple phone manufacturers boast sensitive microphones with high-recording speeds. Accessible technology can be leveraged with novel analytical methods in this era of increasing computational power.

The primary aim of this funded study was to investigate the feasibility of obtaining mobile phone recordings in busy emergency departments with the intention of detecting respiratory infections such as in the setting of new infectious disease outbreaks. A secondary goal was to determine the ability to train predictive models on as few as 50 positive testing cases, which would support rapid response diagnostic counter measures.

Methods

Settings and participants

This prospective study enrolled a convenience sample of patients presenting to one of three emergency departments within one academic health system (a Level I Trauma Center, an urban tertiary care center, and a stand-alone rural ED). Prior to enrollment, stated aims of the overall project were to delineate five groups of subjects: influenza A, pneumonia, acute bronchitis, other respiratory conditions, and controls. Controls were defined as subjects with no acute or chronic respiratory symptoms or diagnoses.

Enrollers approached any ED patient who met inclusion criteria: ED evaluation for any complaint, English speaking, consentable/decisional adults. Personnel used the Fig. 1 algorithm to determine enrollment group. For the primary focus of this paper, pneumonia diagnosis, study personnel used ICD-10 diagnosis codes in combination with clinical signs and imaging results. Because clinical features and imaging factor into the decision to give ICD-10 codes, all imaging, lab and clinical signs contributed to the delineation of pneumonia. The PI reviewed charts and adjudicated all pneumonia cases to confirm the diagnosis; the PI had no role in the the analysis of the recordings.

Fig. 1.

Enrollment algorithm

To minimize interference, the protocol excluded patients requiring immediate stabilization/intervention and those with mechanical heart valves, pacemakers, or implantable cardioverter defibrillators. We also excluded pregnant patients and those under 18 years of age.

Ethics statement

The study was approved by the institutional IRB. Patients were approached after initial ED assessment and urgent clinical tasks, to ensure no delays in care. The study team did not advise or perform any testing or treatment; no subjects had alterations in care based on the study. All subjects signed informed consent prior to recordings and abstraction of clinical data from the electronic medical record.

Data collection and storage

Demographic and clinical data were obtained from the electronic medical record and stored in RedCap, an encrypted research database. Recordings were obtained using a selection of mobile phone brands: iPhone 14, iPhone 15, Nokia C01 Plus (Model TA-1391), and TCL 30 Hz Straight Talk (Model SM-S134DL[GP]). Proprietary recording software was developed with funds from the Biomedical Advanced Research and Development Authority (BARDA).

Unmodified mobile phones were pressed on the microphone side directly against the skin, reducing capture of background noise. The recording process was not compromised by ambient noise in any of the three EDs. Thirty-second recordings were completed in 44.1 khz, WAV format, 16-bit depth which is a recording industry standard and within the capability of most if not all mobile phones. No calibration procedures were necessary to ensure comparability across device types, such as amplitude normalization, bandpass filtering, or device-specific correction factors. Between enrollments, the phones were cleaned with alcohol pads.

The auscultation method included recordings at the right and left anterior axillary line during normal and deep breathing, along with egophony at the right supraclavicular fossa (a technique used to detect Covid-19 in prior work⁹). To recreate a stable and reproducible airflow, egophony involved the patient saying “e” for two full breaths. The axillary recordings utilized normal and deep breaths with no phonation. Sound recording files were uploaded into CardBox (Box, Inc.), a web-based, encrypted could platform. Fig. 2 show screenshots from the recording app that was developed for this study.

Fig. 2.

Recording sites and app screenshots. Source: Tele-stethoscope, Inc

Enrollers carried out a brief oral patient questionnaire and queried the medical record. Subjects were screened with the validated Dyspnea Symptom Severity Scale [5], which asks subjects to categorize their degree of breathlessness into five categories. Other clinical characteristics and outcomes data were abstracted from the medical record. Only IRB-approved study personnel had access to the data. Subjects did not receive incentive for participation.

Technology and analytic method

The mobile phone auscultation (MPA) device comprises a recording application that allows capture of data through standard mobile phone microphones followed by nonlinear dynamics analytical process. The recording application functions like a high-speed voice recorder, supports multiple sites, and provides a waveform guide. In prior work, this device has achieved detection of COVID diagnosis [6], volume overload [7], heart failure/low ejection fraction [8], and chronic obstructive pulmonary disease [9].

The study subcontractor developed diagnostic algorithms using the novel nomenclature of Time Series Dynamics (TSD) modeling software, based on the recordings and de-identified information in REDcap. TSD modeling is a rigorous computational nonlinear dynamics approach to modeling time series observations of low dimensional chaotic systems for the purposes of prediction or classification. When applied to auscultation data, TSD modeling is effectively a chaos-based approach to extracting and fitting features to health states. The purpose of this analysis was to set a performance benchmark for the TSD modeling approach in terms of information that could be known in impromptu examinations of remote patients (e.g. age, gender, symptoms).

Recordings were obtained using unmodified mobile phone hardware equipped with software paired and subsequently analyzed with computational nonlinear biofluid dynamics to classify organ functionality. All recordings were evaluated for Maximal Lyapunov Exponent (MLE) and Correlation Dimension (Dcorr) to confirm the presence of low dimensional chaos and applicability of Time Series Dynamics (TSD) modeling. Train and test sets were created by 80/20 random sampling of clustered records to reduce the risk that both sets were not sampled from the same universe. Features were also tested by Kolmogorov-Smirnov test to help ensure that the test and train sets were drawn from the same universe. Binary classifiers were fitted by using logistic regression. The recordings were evaluated by standard confusion matrix analysis for both the test and train sets.

The candiate features were reduced to two vectors through a combinatorial approach with no fitting or knowledge of the test set. In this approach, which utilizes two degrees of freedom, the feature selection was optimized by various functions such as minimum entropy. Different optimization strategies yielded a few different vector combinations.

Logistic regression was then used to fit the dependent variable to the two vector combinations. Test set data was passed through the fitted model and reported with training set results for sensitivity, specificity, and area under the curve (AUC). Comparison of the test and train set performance serves as a check against overfitting and inconsistent representation. The accuracy of the model was evaluated across various patient subgroups.

Statistical methods

Demographic variables, medical comorbidities, duration of symptoms, Dyspnea symptom severity scale scores, laboratory values (e.g., white blood cell count and lactic acid), and emergency department disposition variables are reported using frequencies and percentages for categorical variables and means with standard deviations or medians with interquartile ranges (25th and 75th percentiles) for continuous variables, where appropriate.

Comparisons on the variables between pneumonia patients and all patient groups combined were conducted using Pearson’s Chi-square test or Fisher’s exact test for categorical variables, the independent-samples t-test for normally distributed continuous variables, and the Mann–Whitney U rank sum test for non-normally distributed continuous variables.

Among the five groups (pneumonia, influenza A, bronchitis, other respiratory disease, and no respiratory disease), categorical variables were analyzed using Pearson’s Chi-square test or the Fisher–Freeman–Halton exact test. If significance was discovered among groups, uncorrected pairwise z-tests were performed to identify between which groups significance was found.

For normally distributed continuous variables, one-way analysis of variance (ANOVA) was performed and if significance was found among groups, uncorrected pairwise post hoc comparisons were conducted between each group to assess group differences. For non-normally distributed continuous variables, the Kruskal–Wallis test or the Kruskal–Wallis exact test was used and if significance was found among groups uncorrected post hoc pairwise comparisons were made between groups using Dunn’s test.

Sensitivity, specificity, positive predictive values (PPV) and negative predictive values (NPV), along with 95% confidence intervals were assessed for xrays finding of pneumonia with actual diagnosis as determined by clinician diagnosis. All statistical tests were two-tailed, and significance was set at p <.05. Analyses were conducted using SPSS for Windows, Version 30.0 (IBM Corp., Armonk, NY) and SAS^® 9.4 (SAS Institute Inc., Cary, NC), specifically using the PROC FREQ procedure to calculate sensitivity, specificity, PPV, and NPV.

Results

From Nov 21, 2023 to June 12, 2024, the team enrolled 292 total subjects with mean age 49 years (SD 17.2), 50.0% male sex; 68.0% white and 29.2% Black. The protocol nearly reached predetermined targets of 65 subjects per cohort: 64 pneumonia, 59 influenza A, 38 acute bronchitis, 68 other respiratory, and 63 controls.

In binary comparison of pneumonia subjects to all other groups, those with pneumonia were older (58 years vs. 46 years), more likely male (56% vs. 48%), more likely white (77% vs. 66%), and more likely to have a tobacco history (77% vs. 68%). See Table 1 for baseline characteristics in the pneumonia group compared to all other groups. Table 2 compares clinical characteristics (symptom duration, dyspnea scale, labs, and disposition) between pneumonia group and all others.

Table 1.

Demographics and medical comorbidities pneumonia vs. other subjects combined

Variable	Total (N = 292)	Pneumonia (n = 64)	Comparison (n = 228)	p*
Age – years, Mean (SD)	48.9 (17.2)	58.0 (13.2)	46.0 (17.3)	< 0.001
Female sex – n (%)	146 (50.0%)	28 (43.8%)	118 (51.8%)	0.258
Race – n (%)				0.577
White	198 (68.0%)	49 (76.6%)	149 (65.6%)
Black/African American	85 (29.2%)	14 (21.9%)	71 (31.3%)
American Indian/ Alaskan	1 (0.3%)	0 (0.0%)	1 (0.4%)
Asian	1 (0.3%)	0 (0.0%)	1 (0.4%)
Other	6 (2.1%)	1 (1.6%)	5 (2.2%)
Ethnicity, Non-Hispanic - n (%)	281 (96.6%)	64 (100.0%)	217 (95.6%)	0.125
Tobacco History – n (%)	205 (70.2%)	49 (76.6%)	156 (68.4%)	0.208
BMI kg/m2 – Median (IQR)	27.90 (23.6, 34.0)	28.70 (23.9, 35.5)	27.8 (23.6, 33.9)	0.629 **
BMI Category – n (%)
< 18.5	12 (4.1%)	2 (3.1%)	10 (4.4%)	1.000
18.5–24.9	87 (29.8%)	21 (32.8%)	66 (28.9%)	0.550
25–29.9	76 (26.0%)	12 (18.8%)	64 (28.1%)	0.133
30–34.9	51 (17.5%)	12 (18.8%)	39 (17.1%)	0.759
>/= 35.0	66 (22.6%)	17 (26.6%)	49 (21.5%)	0.391
Medical comorbidities
0	33 (11.3%)	2 (3.1%)	31 (13.6%)	0.019
1	84 (28.8%)	11 (17.2%)	73 (32.0%)	0.028
2	80 (27.4%)	16 (25.0%)	64 (28.1%)	0.627
3	41 (14.0%)	13 (20.3%)	28 (12.3%)	0.102
4	29 (9.9%)	11 (17.2%)	18 (7.9%)	0.028
5+	25 (8.6%)	11 (17.2%)	14 (6.1%)	0.005

^*Parametric testing performed using independent-sample t test, chi-squared or Fischer’s exact where appropriate based on cell size appropriate

^**Non-parametric testing performed using the Mann-Whitney rank sum test

Table 2.

Clinical characteristics for all subjects stratified by sub-group

	Total (N = 292)	Pneumonia (n = 64)	Comparison (n = 228)	p*
Duration of symptoms – n (%)
< 1 day	42 (14.9%)	1 (1.6%)	41 (18.8%)	< 0.001
1–2 days	59 (20.9%)	12 (18.8%)	47 (21.6%)	0.827
3–4 days	64 (22.7%)	17 (26.6%)	47 (21.6%)	0.401
5–7 days	41 (14.5%)	10 (15.6%)	31 (14.2%)	0.779
> 7 days	76 (27.0%)	24 (37.5%)	52 (23.9%)	0.031
Dyspnea Symptom Severity Scale – n (%)
0	79 (27.2%)	10 (15.6%)	69 (30.5%)	0.018
1	64 (22.1%)	7 (10.9%)	57 (25.2%)	0.015
2	39 (13.4%)	10 (15.6%)	29 (12.8%)	0.563
3	76 (26.2%)	25 (39.1%)	51 (22.6%)	0.008
4	32 (11.0%)	12 (18.8%)	20 (8.8%)	0.026
Dyspnea Symptom Severity Scale – Median (IQR), n	2 (2, 2), 290	3 (1,3), 64	1 (0, 3), 226	< 0.001**
WBC – Median (IQR), n	8.20 (6.50, 10.20)	9.45 (6.30, 12.70)	7.60 (6.90, 9.40)	0.213 **
Lactic Acid – Median (IQR), n	1.40 (0.90, 2.00)	1.37 (0.90, 2.00)	1.55 (0.97, 2.10)	0.482 **
ED Disposition – n (%)
Admit to Floor/PCU	104 (35.6%)	42 (65.6%)	62 (27.2%)	< 0.001
Admit to ICU	3 (1.0%)	1 (1.6%)	2 (0.9%)	0.631
Discharged Home	170 (58.2%)	13 (20.3%)	157 (68.9%)	< 0.001
AMA	2 (0.7%)	1 (1.6%)	1 (0.4%)	0.391
Death	1 (0.3%)	1 (1.6%)	0 (0.0%)	0.219
Other	12 (4.1%)	6 (9.4%)	6 (2.6%)	0.027
Length of stay (in hours) – Median (IQR), n	5.50 (2.72, 28.23	19.43 (7.93, 96.72)	4.13 (2.45, 23.92)	< 0.001**

^*Parametric testing performed using chi-squared or Fischer’s exact where appropriate based on cell size appropriate

^**Non-parametric testing performed using the Mann-Whitney rank sum test

Analysis of MPA recordings

No completed recordings were excluded from analysis. Features were extracted from the recordings under the general assumption of chaos using extensive but common time series and phase space methodologies. Selected features were then combined, without weighting or fitting, into a few different vector solutions. The selection was made by metaheuristic solvation with a few different optimization functions. The most fruitful of these was minimum entropy which corresponds roughly to maximum variance but without the linear constraints of primary component analysis. Some experimentation with logistic regression allowed a two-vector solution which is well within Peduzzi simulation guidance.

TSD models were initially fitted separately to the axillary and egophony recordings and then combined into a “composite” model. The most accurate results were found with egophonic, right normal breathing, and composite of those two. Evaluation of the egophony train set recordings shows good results at low pass of 450 hz. There was also a meaningful peak at 50 hz in the egophony recordings that was likewise observed in the right lung. Fig. 3 illustrates raw sound in a patient positive for pneumonia.

Fig. 3.

Raw sound visual representation for pneumonia subject. Source: Fleming Scientific

The models yielded one false negative and seven false positives. The right normal breathing (sens 85%, spec 81%) and egophonic (sens 85%, spec 86%) models performed well. The composite model achieved sensitivity 91%, specificity 87%, and area under the curve of 89% (Table 3). Table 4 illustrates diagnostic performace of Chest X-Ray in the overall cohort. The accuracy of the model was evaluated across various patient groups. In general, the model performed consistently across different demographics, BMI, and tobacco use history (Table 5).

Table 3.

Modeling of auscultation recordings

Metric	Composite	Right Normal Breathing	Egophony
AUC-Train	0.893	0.839	0.865
AUC-Test	0.885	0.814	0.847
AUC-Sum	1.779	1.654	1.712
Sensitivity-Train	0.902	0.843	0.863
Sensitivity-Test	0.923	0.846	0.846
Sensitivity -Avg	0.913	0.845	0.854
Specificity-Train	0.885	0.835	0.868
Specificity-Test	0.848	0.783	0.848
Specificity -Avg	0.866	0.809	0.858
True Negative-Train	161	152	158
False Negative-Train	5	8	7
True Positive-Train	46	43	44
False Positive-Train	21	30	24
Total Cases-Train	233	233	233
True Negative-Test	39	36	39
False Negative-Test	1	2	2
True Positive-Test	12	11	11
False Positive-Test	7	10	7
Total Cases-Test	59	59	59
Total Cases	292	292	292

Table 4.

Chest x-ray (CXR) performance for pneumonia diagnosis

	Pneumonia Diagnosis	No Pneumonia Diagnosis
Pneumonia on CXR	52	56
No Pneumonia on CXR	8	93
CXR Sensitivity	87%	(78%, 95%)*
CXR Specificity	62%	(55%, 70%)*
CXR Positive Predictive Value	48%	(39%, 58%)*
CXR Negative Predictive Value	92%	(87%, 97%)*

^*95% Confidence Limits

^**Four subjects in the pneumonia group did not have a CXR

Table 5.

TSD modeling accuracy across patient subgroups

Group (n=)	Accuracy (%)
Overall (n = 292)	88.4%
Age > 49 (n = 143)	85.3%
Age ≤ 49 (n = 149)	91.3%
Male (n = 146)	87.7%
Female (n = 146)	89.0%
White (n = 198)	88.9%
Non-White (n = 94)	87.2%
BMI > 30 (n = 118)	88.1%
BMI ≤ 30 (n = 174)	88.5%
Tobacco Never (n = 87)	86.2%
Tobacco Ever (n = 205)	89.3%

Discussion

This study achieved the primary aim of establishing feasibility to accurately diagnose pneumonia using mobile phone recordings in small enrollment groups. In a diverse cohort with varied respiratory conditions, models of right normal breathing, egophony, and their composite achieved notable sensitivity, specificity, and AUC. This study was funded by BARDA as a medical countermeasures study, with intention to train a diagnostic technology on small samples of patients (as few as 50) verified to have an acute medical condition (such as infection with a new pathogen).

A study in more than two million admitted patients found that 57% had discordance in initial versus discharge diagnosis. Patients who lacked an initial pneumonia diagnosis had a significantly higher 30-day mortality rate [10]. A study of 17,290 patients found that up to 12% of subjects hospitalized with CAP met criteria for misdiagnosis [11]. While lung ultrasound offers improved detection compared to radiographs [12, 13], current diagnostic approaches (clinical or imaging) continue to lack sensitivity and specificity.

Pneumonia diagnostic methods can be classified into four categories: laboratory-based, acoustic- or chest-sound-based, imaging-based, and physiological-measurement-based. Stethoscope auscultation represents the normal standard to diagnose pneumonia, but has limitations such as experience and expertise. Chest radiography suffers from poor sensitivity and specificity. While more sensitive and specific, chest CT involves more specialized equipment, higher radiation exposure/cost, and inability to scale to larger populations.

The technology in this study concerns the 2nd category, auscultatory methods that “involve using digital devices such as microphones to pick up the acoustic signals from the chest” [14]. These techniques can have reasonable accuracy and sensitivity, but often lack specificity. Most remain in prototype phase and are not yet commercially available [14]. These models used recordings from egophony at the right supraclavicular fossa and normal breathing at the right axillary site. Egophony provides a method of stabilizing airflow to standardize recordings. From the recordings, the sponsor extracted and analyzed data to characterize biofluid dynamics (true physical sounds) without machine learning.

Pneumodynamics result in energy that is dissipated to the surface of the skin where it can be detected as vibration. In traditional auscultation, a transducer converts this vibration into audible sound by creating a pressure change in trapped air. By trapping air between the phone microphones and patient’s skin, a similar mechanical process takes place. However, the volume is too small to produce audible sound. Instead, the information is passed to a time series (aka waveform) of changing energy intensity measured as decibels. The way that systems dissipate energy over time is, by definition, “turbulence.” Turbulence is the physical manifestation of chaos. Consequently, the raw data captured by the phones is pneumodynamic turbulence and it is classified in terms of chaos-based markers. The approach is unrelated to frequency domain or other acoustic signal analysis processes and would not be relatable to crackles, rails etc.

Kanwal et al. reviewed several key studies related to use of auscultatory diagnosis of pneumonia over the past ten years. The devices capture chest sounds, cough sounds, or other respiratory sounds [15–20]. The technologies often incorporate some degree of machine learning or computer analysis of the sounds obtained by devices such as phones or other proprietary hardware.

Other diagnostic devices have integrate stethoscope recording in attempts to diagnose COVID-19 pneumonia [21, 22]. Gheisari et al. summarized 42 papers presenting mobile apps intended to diagnose COVID-19. The majority of these apps utilized contact tracing, symptoms, etc., rather than measuring or recording data from participants (14 used sound and voice, three used image and sound) [23]. Compared to COVID-19 diagnosis and detection, relatively few publications have presented novel technology to diagnose bacterial pneumonia as in this study.

One study used self-recording by patients on their suprasternal notch to delineate COPD patients with pneumonia from those simply experiencing an exacerbation. The approach used hybrid system based on principal component analysis (PCA) and probabilistic neural networks (PNNs) [24]. Other research has attempted to use voice analysis to detect respiratory disease, though efforts seems to only focus on COVID-19 [25].

The increasing prevalence of telemedicine would benefit from a scalable technology. A non-invasive, cost-effective, readily available, and skill-independent test could bring respiratory diagnostics to underserved areas worldwide. Recordings with mobile phones’ built-in microphones represents a promising technology for monitoring respiratory conditions outside of medical facilities. Real-time analysis of respiratory sounds’ specific signal characteristics, including biofluid dynamics and turbulence, would allow clinicians to amplify, review, and compare recordings.

Other functions of mobile phones, such as contact tracing and symptom monitoring, could be combined with the phones auscultation potential to allow for a integrated platform to track and manage pandemics. BARDA’s mission is to develop medical countermeasures (MCMs) and increase preparedness for health security threats, with special focus on EIDs. After supporting over 90 MCMs through FDA approval, BARDA plans to implement a new three step strategy in 2025: “ [1] prioritize and partner [2], develop and sustain, and [3] validate capabilities for the most rapid response to the widest variety of EIDs.” [26].

The technology described here intends to serve as an adjunct for diagnosis and Rapid Response to new threats. We found no barriers to obtaining recordings with unmodified phones, no limitations in background noise, and easy decontamination of devices between patients. With the ability to develop new models with as few as 50 cases, the device could be used in population health situations. Settings with no radiographic capability would be intuitive locations to deploy this type of technology. While we found few false positive and false negative results, future studies in larger patient populations would be necessary to deploy the technology for clinical decision making. Pretest probability and Bayesian considerations would also apply.

Limitations

This study has several limitations. While diverse, the relatively small cohort was derived from three EDs in one city in the Southeast US. The approach was not intended to develop robust diagnostic models; instead we aimed to simulate incursion of a new threat and the ability of experimental software to create rapid response, or medical counter measures, solutions. A formal approach would utilize more positive cases, making validation modeling possible.

While this study aimed for feasibility and not diagnostic accuracy, we have included clinical data on the one false negative and seven false positive cases. The single false negative subject was a 55-year-old female nonsmoker with no past medical history who presented with 5–7 days of infectious symptoms; chest xray showed lobar consolidation at the right lung base. The seven false-positive cases had a mix of smoking status, gender, BMI, and other factors: chronic lung disease, cardiopulmonary comorbidity, and radiographic ambiguity (e.g., atelectasis or interstitial changes rather than infectious consolidation). The seven subjects fell into each of the other enrollment groups. The imbalance between false positives and false negatives suggests that the technology may preferentially err toward higher sensitivity than specificity, a characteristic that may be desirable in screening or early-warning contexts. As discussed elsewhere, the aim of this study was feasibility and not diagnostic accuracy.

Results had no significant accuracy drop off in subgroups, but a larger study is needed to draw more definitive conclusions about limitations posed by demographics, body type and comorbidities. As the device is based on nonlinear dynamics, and not frequency domain, diagnostic confusion can only occur when dynamics are similar. The presence of abnormal lung sounds can only impact the accuracy of the device if it is associated with identical turbulence patterns but not if it is associated only with frequency domain interference.

Patients with implanted pacemakers and cardioverter/defibrillators were excluded due to concern from regulatory agencies for the small potential for interaction of the magnets in mobile phones with these devices. Research on older phone models suggests that mobile phones do not cause interference unless placed in “very close proximity” to these devices [27]. The microphone is on the narrow end of the phone, thus the back of the phone with the magnet did not touch the patient’s skin. Of note, pacemakers and ICDs should not interfere with the acoustic recordings or our analysis but we excluded out of abundance of caution for potential for harm to patients.

The results were evaluated by standard confusion matrix analysis for both the test and train sets. Comparison of the test and train set performance serves as a check against overfitting and non-uniform representation of the underlying universe. Examination of common spectral markers, such as those from power spectrum, showed no modeling utility. All of the surviving features came from the nonlinear dynamics group which was expected given the chaotic nature of the signals.

As common acoustic analysis typically assumes limit cycle attraction and/or linear phasal relationships, these features were not as stable or robust. In a larger phase 2 study the training step would utilize a validation process; but the data was too limited for this study and the authors believe that the shortened process was adequate for establishing feasibility.

Future studies will deploy the technology to larger cohorts in Phase 2 protocols in remote settings, telehealth, EDs, and ambulatory care settings. As there were relatively few positive cases, it as not possible to perform a true test/train analysis with a validation step. Clinical studies will ask patients to obtain recordings, proving utility of the technology outside of medical facilities. Those studies will be larger (for statistical significance) and recruit to obtain population/demographic representation across multiple sites. Other validation steps (beyond population demographics) include safety and effectiveness studies required by the FDA, quality assurance algorithms to ensure recordings were properly made and transmitted, cyber security and HIPPA and other International Organization for Standardization 13,485 requirements.

Conclusion

This funded research found that chaos-based, time series dynamics modeling of data from unmodified mobile phone recordings yielded excellent sensitivity, specificity, and area under the curve to delineate pneumonia from other respiratory illness and control subjects. This technology could increase patient access to accurate diagnosis of multiple medical conditions, particularly in the setting of medical counter measures for disease outbreaks, as intended in this study. Phase 2 will further characterize these models for deployment in larger, more diverse populations.

Abbreviations

AUC

Area Under the Curve

BARDA

Biomedical Advanced Research and Development Authority

BMI

Body Mass Index

CAP

Community-Acquired Pneumonia

COPD

Chronic Obstructive Pulmonary Disease

CXR

Chest X-ray

Dcorr

Correlation Dimension

ED

Emergency Department

EID

Emerging Infectious Disease

FDA

Food and Drug Administration

HIPAA

Health Insurance Portability and Accountability Act

ICD-10

International Classification of Diseases, Tenth Revision

MLE

Maximal Lyapunov Exponent

MPA

Mobile Phone Auscultation

PCA

Principal Component Analysis

PNN

Probabilistic Neural Network

TSD

Time Series Dynamics

WAV

Waveform Audio File Format

Authors’ contributions

Conceptualization: MHMethodology: MH, HS, JT, JMInvestigation: MH, HS, JMFormal analysis: CZ, RCData curation: CZ, RCWriting – original draft: MH, HS, JMWriting – review & editing: MH, JT, CZ, RCSupervision: MH, JT, RCProject administration: MH, HS, JMFunding acquisition: MH.

Funding

This project has been supported in whole or in part with federal funds from the Department of Health and Human Services; Administration for Strategic Preparedness and Response; Biomedical Advanced Research and Development Authority (BARDA).

Data availability

The datasets generated and/or analyzed during the current study are not publicly available due to privacy considerations and institutional restrictions related to protected health information, but some of the de-identified data may be made available from the corresponding author on reasonable request, subject to approval by the University of Louisville, Fleming Scientific, TSI, and BARDA.

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the principles of the Declaration of Helsinki. The study protocol was reviewed and approved by the University of Louisville Institutional Review Board (IRB). All participants provided written informed consent prior to enrollment, audio recordings, and abstraction of clinical data from the electronic medical record.

Consent for publication

As above, covered by UofL IRB.

Competing interests

The authors declare no competing interests.