Development of an Automated Ultrasound Signal Indicator of Lung Interstitial Syndrome

Abstract

Objectives:

The number and distribution of lung ultrasound (LUS) imaging artifacts termed B-lines correlate with the presence of acute lung disease such as infection, acute respiratory distress syndrome (ARDS), and pulmonary edema. Detection and interpretation of B-lines requires dedicated training and is machine and operator dependent. The goal of this study was to identify radio frequency (RF) signal features associated with B-lines in a cohort of patients with cardiogenic pulmonary edema. A quantitative signal indicator could then be used in a single-element, non-imaging, wearable, automated lung ultrasound sensor (LUSS) for continuous hands-free monitoring of lung fluid.

Methods:

In this prospective study a 10-zone LUS exam was performed in 16 participants, including 12 patients admitted with acute cardiogenic pulmonary edema (mean age 60±12 years) and 4 healthy controls (mean age 44±21). Overall,160 individual LUS video clips were recorded. The LUS exams were performed with a phased array probe driven by an open platform ultrasound system with simultaneous RF signal collection. RF data were analyzed offline for candidate B-line indicators based on signal amplitude, temporal variability and frequency spectrum; blinded independent review of LUS images for presence or absence of B-lines served as ground truth. Predictive performance of the signal indicators was determined with receiving operator characteristic (ROC) analysis with k-fold cross validation.

Results:

Two RF signal features – temporal variability of signal amplitude at large depths and at the pleural line – were strongly associated with B-line presence. The sensitivity and specificity of a combinatorial indicator were 93.2% and 58.5%, respectively, with cross-validated area under the ROC curve (AUC) of 0.91 (95% CI = 0.80-0.94).

Conclusion:

A combinatorial signal indicator for use with single element non-imaging LUSS was developed to facilitate continuous monitoring of lung fluid in patients with respiratory illness.

Introduction

Emergency Department (ED) patients who present with acute respiratory illness must be triaged quickly, accurately diagnosed, and monitored for respiratory failure throughout hospitalization. Chest x-ray and CT are typically used for this purpose, but cannot be done continuously or serially and are associated with logistical limitations, for example when transporting unstable patients with hypoxemia or patients with respiratory infection due to the risk of contagion. Ultrasound is non-ionizing, rapid and accessible and has been shown to have high sensitivity for the diagnosis of pneumonia (including COVID-19), pulmonary edema, and ARDS^1-7. Lung ultrasound (LUS) can also be used to monitor progression or improvement of disease and to adjust treatment regimen^8-12. Specifically, LUS imaging artifacts — B-lines — are known to be correlated with the presence of lung interstitial syndrome and severity of the disease.^9,13-15 Per standard LUS protocol, scanning is performed in 6-10 anatomic zones to interrogate different locations of the lung, and the number and distribution (focal versus diffuse) of B-lines in each zone are determined. B-lines represent acoustic reverberations within regions of alveolar or interstitial edema adjacent to the pleural line, but the exact mechanisms by which they form are still not fully understood^{4, 16-17}.

One limitation of LUS is that the detection and quantification of B-lines requires specific training and is machine and operator dependent¹⁸. While a number of recent studies on machine learning-based algorithms for automated B-line detection have a potential to address this problem, LUS exam still requires a multi-element US probe, an imaging system, and an operator.¹⁹ Therefore, continuous or frequent monitoring of lung condition through conventional LUS is logistically challenging. Further, in a number of clinical scenarios, it would be preferable to use adhesive wearable lung ultrasound sensors (LUSS) that could be attached independently onto the anatomic zones of interest and interrogated on demand. In order to make such LUSS miniature (approximate size similar to an EKG lead), the pulse-echo and signal processing electronics have to be portable and affordable, which precludes the use of multi-element arrays and image reconstruction software. To satisfy the above requirements, we rationalize that B-lines are artifacts, as opposed to true images, and thus the algorithms and devices to collect data for an image, reconstruct the image, and then interpret it are fundamentally redundant. Instead, we hypothesize that there are ultrasound signal features that are uniquely associated with B-lines in LUS images, and those features can be automatically detected directly from the signals.

The aim of this study was to identify the radio frequency (RF) signal features and derive an automated indicator of B-line presence in a cohort of patients with cardiogenic pulmonary edema. The indicator could then be used in a single-element LUSS for continuous hands-free monitoring of lung fluid.

Materials and Methods

Study Design

This was a prospective observational cohort study conducted at the University of Washington (UW) Medical Center which serves as a regional referral center for advanced cardiac care. The study was approved by UW Institutional Review Board (IRB) and performed in compliance with HIPAA with written participant informed consent. Patients admitted to the hospital with a presumed diagnosis of cardiogenic pulmonary edema were eligible for enrollment, with target enrollment of n=15. Five healthy volunteers were also enrolled from the study team. Excluded were participants who were less than 18 years of age, incarcerated, pregnant, unable to provide consent in English, had known COVID infection, or patients whose condition required acute intervention such as urgent angiography or emergency mechanical device implantation for cardiogenic shock.

Lung Ultrasound Exam and Data Collection

Each participant received a LUS exam per a standard 10-zone protocol illustrated in Fig.1a performed by a certified sonographer or an attending emergency physician trained in lung ultrasound. The protocol consisted of positioning an US transducer (64-element phased array S4-2M, Humanscan, Seoul, South Korea) consecutively over each zone, with the imaging plane perpendicular to the intercostal spaces. With the probe immobile and the subject respirating freely, a 3-second cine loop of B-mode ultrasound was recorded at 15 frames per second (fps), resulting in a 45-frame video per lung zone. The US probe was operated by an open platform ultrasound system - Verasonics Ultrasound Engine V1 (Verasonics, Kirkland, WA, USA) with imaging settings replicating those of the lung preset of Mindray TE7 (Mindray Medical International, Shenzen, China): operating frequency 4.5 MHz, sector angle 45°, number of scan lines 128, focusing depth 2.5 cm, depth 12 cm, MI 0.8.

Figure 1.

Lung ultrasound (LUS) exam and radiofrequency (RF) data collection and processing. (a) Illustration of the 10 anatomic locations to be scanned per LUS protocol. (b) Flow diagram of LUS video capture and collection of corresponding RF signals, $s(t)$, where $t$ is time. B-mode US image was displayed at 15 frames per second (fps), and RF signals for the focused and wide beam acquisitions corresponding to each frame were saved in the background. (c) Illustration of signal processing for a focused beam acquisition to obtain cumulative amplitude, ${CA}_F$, at large depth. ${\mathrm{s}}_{ij}(d)$ – RF signal detected by i^th element at j^th scan line, $d=c_{0}t∕2$ is depth, $c_0$ – speed of sound in tissue.

During the recording of LUS video in each zone, the echoes received by all elements of the array (referred to as radiofrequency (RF) signals) in each frame corresponding to all the scan lines (128 per frame) were recorded at the sampling frequency of 20 MHz. Each set of 128 scan lines forming a LUS image frame was followed by 35 additional US pulses emitted without focusing (i.e. as plane waves in a wide beam) at pulse repetition frequency (PRF) of 5 kHz, and the corresponding RF signals were also recorded at the same sampling frequency and analyzed in post-processing as described below.

Blinded LUS video review

Three emergency department physicians trained in LUS and blinded to the subject cohort (but not the zone number) independently reviewed the LUS videos and provided a binary score indicating presence or absence of B-lines. Cohen’s Kappa was calculated to assess the degree of agreement between two expert clinical reviewers, and a third expert clinical reviewer served as a tiebreaker in cases with discordant image interpretations.

RF signal analysis and candidate indicators of B-lines

RF signal analysis was performed independently of the corresponding LUS video review and is illustrated in Fig.2. The following groups of RF signal metrics were considered as candidates for B-line indicators: signal amplitude at large (8-12 cm) depth, frequency spectrum of the signal, and signal phase and amplitude temporal variability. All of the indicators were computed for each RF signal set corresponding to a single frame in a LUS video, and then maximum, minimum and average over the 45 frames were calculated. The rationale behind selecting those groups of indicators was based on qualities of B-lines known from the literature and observed in practice, and the hypothesized mechanisms of B-line formation.^16,17,20 Specifically, RF signal amplitude at a certain time point determines the brightness of the US image at the corresponding spatial location. Because B-lines are bright artifacts, it is reasonable to hypothesize that they will correspond to RF signals with large amplitude. Furthermore, B-lines by definition extend from the position of the pleural line to at least 12 cm depth²⁰, where no other bright reflectors are expected. To account for differences in body habitus and thus depth of pleural line location, the range of RF signal arrival time, $t$, and corresponding depth, $d=c_{0}t∕2$, where $c_0$ is the speed of sound, was selected as 8-12 cm for signal amplitude analysis. First, cumulative RF signal $S(d)$ was obtained by summation of RF signals $s(d)$ over the central 16 elements of the US probe (for both focused and wide beam acquisitions) and over 32 central scan lines (for the focused acquisitions). Cumulative signal amplitude for both wide beam plane wave and focused beam acquisitions, ${CA}_W$ and ${CA}_F$ respectively, was then calculated by integration of the absolute value of $S(d)$ over depth within the 8-12 cm range. Both of those indicators were then normalized by their corresponding noise values calculated from reference blank LUS acquisitions, where the probe was not in contact with a subject. Mean and maximum $CA$ indicators were computed over 45 frames of each video.

Figure 2.

Representative examples of LUS frames from the same video (Supplementary Video 1) with and without a B-line obtained with focused scan lines ((b) and (a)) and wide beam ((d) and (c)) acquisitions and corresponding RF signals ((c) and (f)). B-line is present in both images (b) and (e), although with lower contrast in the wide beam case, corresponding to lower RF signal amplitude at large depth.

Analyzing the temporal variability of RF signal amplitude at the time scale corresponding to the frame rate of LUS was motivated by the “shimmering” quality of B-lines relatively to echo texture of soft tissues, reverberations, and clutter that may be encountered in LUS. Thus, cumulative signal amplitude variability, $\Delta {CA}_F$, was calculated for focused beam RF datasets corresponding to a video clip as follows:

$$\Delta {CA}_F=({CA}_{Fmax}-{CA}_{Fmin})\cdot 100\%∕{CA}_{Fmean}$$ (1)

where ${CA}_{Fmax}$, ${CA}_{Fmin}$ and ${CA}_{Fmean}$ represent maximum, minimum and mean values of ${CA}_F$ over the 45 frames.

Analysis of the temporal variability of the RF signal at the shorter time scale – within milliseconds – was motivated by the mechanism hypothesized in the literature to be responsible for B-line formation – reverberation of the LUS imaging pulses within irregularly-shaped fluid-filled areas of the lung adjacent to the pleural line.^4,17 The dynamics of this reverberation determines the shape of the RF signal over its arrival time, and is expected to be sensitive to the relative LUS beam position and angle of incidence that both change with physiological motion. Therefore, we hypothesized that RF signal corresponding to a B-line may be highly variable over short – millisecond - time scales, i.e. from one wide beam plane wave pulse to the next. The variability was quantified by Doppler-like processing of the 35 consecutive RF signals from plane wave acquisitions. The RF signals on the central 16 elements were summed, transformed into in-phase and quadrature (I/Q) components through Hilbert transform, and wall filtered by linear regression. Doppler power distribution over depth DP(d) was then computed as autocorrelation with a lag of 1 per Kasai et al.²¹ Doppler threshold was set as the average of DP(d) multiplied by a factor of two, and the maximum value of DP(d) divided by this threshold was the output indicator for this frame, DP. Mean and maximum DP were computed over 45 frames of each video.

Lastly, frequency spectra of the RF signals beamformed and summed over the central 16 elements were computed for a limited number (n=10) of central scan lines that contained well defined B-lines and compared to those that did not contain it. This limited investigation was motivated by a hypothesis posed in a numerical study¹⁶ that the signal frequency spectrum could be indicative of the native frequencies of acoustic traps — fluid-filled areas in the lung — and would thus be informative and indicative of B-lines. However, this limited initial analysis did not reveal any consistent changes in the spectral width or shape associated with B-lines, consistently with Kameda et al¹⁸, and thus further analysis was not pursued.

All of the candidate indicators of B-lines were calculated for all LUS datasets and assigned to two groups — B-lines present or absent — according to the binary outcome of the expert review. Two indicators with the highest level of statistical significance (per two-sided Student’s t-test) and the largest effect size were selected for further analysis of predictive accuracy.

Statistical Analysis of the predictive accuracy of candidate B-line indicators

The binary evaluation of the LUS videos by independent reviewers described above (B-line present/absent) served as the ground truth for evaluating predictive accuracy of the candidate signal indicators. The relationship between the two candidate signal indicators and presence of B-lines was modeled via logistic regression. Predictive performance of the regression model was assessed via k-fold cross-validation (k=10), whereby the total dataset was randomly divided into k groups. The first test data set (used to test the accuracy of the predictive model) was set aside and the remaining (k-1) groups served as the training data set (used to fit the predictive logistic regression model). The process was repeated until all k groups had been used in turn as test and training datasets. Receiver Operator Characteristics (ROC) analysis was used to quantify overall performance of the candidate indicators in terms of cross-validated area under the curve (cvAUC; stata package cvauroc).²² cvAUC was calculated by averaging AUCs from each fold and applying bootstrapping procedures to calculate 95% confidence intervals. Contingency tables were also generated to report classifications and evaluate diagnostic accuracy, sensitivity, specificity, positive predictive value, and negative predictive value. Stata 15.1 (College Station, TX) was used for all statistical analyses.

Results

LUS data collection and blind video review.

From May to November 2021, 157 participants were screened, 29 were eligible, and 12 were enrolled with complete imaging. Reasons for non-enrollment of eligible participants included staff being unavailable (9), participants not consenting (4) or other (3). Ultimately 12 patients and 5 healthy volunteers were scanned using the 10-zone LUS protocol; the video clips were not collected for one healthy volunteer due to a technical issue with data recording, and thus were not included in the study. Patients enrolled in the study were admitted from the emergency department for a presumptive diagnosis of acute decompensated heart failure, with all 12 having chest radiograph findings consistent with acute pulmonary edema. Patient demographics, pertinent diagnostic results, and dispositions are illustrated in Table 1.

Table 1.

Patient demographics

Subject	Age	Sex	Past Medical History	Presenting Complaint	Chest Xray Interpretation	Days from admit to LUS	Discharge Diagnosis	Hospital disposition
1	62	Male	None	Cardiogenic shock, heart failure	Acute pulmonary edema. No pleural effusions.	5	Acute decompensated heart failure	Discharged
2	33	Male	Lung disease,CHF	Acute decompensated heart failure	Diffuse lung disease consistent with pulmonary edema.	4	Cardiogenic shock	Discharged
3	90	Female	Lung disease,CHF,CAD	SOB, chest pain	Pulmonary edema with bilateral pleural effusions.	2	CHF	Discharged
4	41	Male	CHF	SOB	Increased lung markings consistent with pulmonary edema.	1	CHF exacerbation	Discharged
5	57	Male	None	Flash pulmonary edema and heart failure	Mild pulmonary edema	1	Acute decompensated heart failure	Discharged
6	59	Male	Lung disease,CHF	SOB, chest pain	Diffuse lung disease concerning for pulmonary edema.	0	Hypertensive urgency	Discharged
7	67	Female	Lung disease,CAD	Flash pulmonary edema	Diffuse lung disease, consistent with edema. Bilateral pleural effusions.	2	Acute heart failure	Died in hospital
8	73	Male	Lung disease, CAD	SOB	Diffuse lung disease consistent with edema and small left pleural effusion.	1	CHF, CAD	Discharged
9	54	Female	CAD	SOB, leg swelling	Increased pulmonary edema. Similar bilateral pleural effusions.	1	Chronic systolic heart failure	Discharged
10	67	Male	Lung disease,CHF,CAD	SOB, leg swelling	Mild bilateral interstitial pulmonary edema. New left pleural effusion.	1	CHF exacerbation	Discharged
11	62	Male	Lung disease,CHF	SOB, leg swelling	Pulmonary edema. Right pleural effusion.	0	Heart failure	Discharged
12	58	Male	Lung disease,CHF	SOB, chest pain	Prominent lung interstitium related to edema. Bilateral pleural effusions.	0	Heart failure	Discharged

CAD- coronary artery disease, CHF- congestive heart failure, SOB - shortness of breath

Overall,160 individual LUS video clips were recorded and scored by 2 blinded reviewers for the presence or absence of B-lines. Inter-rater agreement was 89%, kappa 0.79 (95% CI = 0.63-0.94). Discordant interpretations were adjudicated by consensus with the third expert reviewer in 10.2% of cases, resulting in 77 and 83 video clips with B-lines absent and present, respectively. Complete sets of quality RF signals were collected for 114 video clips, of which 41 and 73 corresponded to the LUS videos with B-lines absent and present, respectively. Those datasets were included in the statistical analysis of candidate B-line indicators; the excluded datasets were either incomplete (n=30 from three healthy volunteers and n=6 from patients, due to data recording issues) or contained noisy RF data (n=10 from a patient with very thick chest wall) such that not all of the candidate metrics could be computed.

RF signal analysis and selection of candidate B-line indicators

Representative examples of a LUS image (from Supplementary Video 1) with and without a B-line reconstructed from consecutively acquired focused scan lines and wide beam, with corresponding RF signals, are shown in Fig.2. B-lines formed in LUS images in the case of wide beam acquisitions, albeit with diminished contrast vs focused scan line images. This is contrary to previously reported expectations that a narrow LUS beam is required for trapping of LUS pulses and B-line formation.¹⁶ The brightness of B-line in LUS images corresponded to higher RF signal amplitude at large depth for both focused scan lines and plane wave acquisitions, although in the latter case the RF signal dropped below the noise floor at depths beyond 10 cm. The lower RF signal amplitude was to be expected, as the pressure in the wide LUS beam incident onto the lung surface was much lower than in the focused scan lines. To account for this, the depth range of 6-10 cm was chosen for batch analysis of cumulative amplitude ${CA}_W$ in wide beam RF signals.

LUS dataset examples illustrating the temporal variability of the RF signal amplitude are presented in Fig.3 and Supplementary Videos. Variability at the short time scales, described by $DP(z)$, typically had low amplitude peaks, often below noise floor, within the depth range corresponding to the chest wall, as expected for perfused tissue. In datasets without B-lines (e.g. Fig.3a, Supplementary Video 2) $DP(z)$ was most commonly at the noise floor throughout depth, with occasional small amplitude peaks within the chest wall and/or at the pleural line. A much higher amplitude peak was consistently observed at the position of the pleural line only in the cases when B-lines were present (Fig.3c). The amplitude of the peak was qualitatively correlated with the number and brightness of B-lines, as well as the amount of their motion with respiration. In most cases this peak was narrow and confined to the location of the pleural line, with $DP(d)$ returning to the noise floor at larger depths. However, there were instances, typically associated with very bright, pronounced B-lines (e.g. Fig.2b), where the peak was wider and extended up to 2 cm depth below the pleural line. The observations above and analysis of the corresponding sets of RF signals led to the understanding that respiratory sliding motion of the lung surface relatively to the LUS beam results in the formation of the $DP$ peaks at the pleural line. As illustrated in Fig.3d, when fluid-filled areas of the lung that result in formation of B-lines are moving across the LUS beam, the amplitude and phase of the wave reflected from the pleural line will change from one LUS pulse to the next, thus producing a peak in $DP$. Conversely, if the lung is fully aerated, its surface is uniformly reflective, and thus lung sliding relatively to LUS beam does not cause measurable changes in reflectance and associated $DP$ peak.

Figure 3.

Representative examples of LUS frames from the videos without B-lines (Supplementary Video 2) (a) and with multiple stationary B-lines(Supplementary Video 3) (b). (c) Distribution of $DP$ over depth for acquisitions corresponding to (a) (blue line), (b) (black line) and Fig.2(b) (red line). The presence of B-lines was typically associated with a pronounced peak in $DP$ co-located with the position of the pleural line, with the peak amplitude generally larger for mobile B-lines than for stationary ones. (d) Illustration of the mechanism for the formation of $DP$ peak at the pleural line: fluid-filled areas of the lung adjacent to the pleural line slide across the LUS beam due to respiration. The corresponding changes in reflectance of the lung surface encountered by subsequent pulses in the Doppler ensemble result in formation of the $DP$ peak. (e), (f) Changes in ${CA}_F$ and peak $DP$, respectively, across the frame number for the same LUS videos as used in (c).

Signal amplitude variability at the longer time scale was characterized by changes in ${CA}_F$ across the frames; representative examples of those changes for the datasets with no B-lines, stationary B-lines and mobile B-lines are presented in Fig.3e. As seen, even in the case of relatively stationary B-lines the variability is larger than the case without B-lines, where signal amplitude consists of A-lines, reverberations and other clutter. The changes of $DP$ over the frames for the same three cases are shown in Fig.3f. It can be appreciated in the case of a mobile and transient B-line that the maximum in $DP$ either precedes or follows the maximum of ${CA}_F$, but does not coincide with it, further supporting the observation that the peak in $DP(d)$ at the pleural line forms due to the sliding of fluid-filled areas (and thus the B-lines) in and out of the LUS beam.

The results of the initial cumulative analysis of candidate B-line indicators are shown in Figs.4(a). All candidate indicators were statistically different between datasets with and without B-lines, although with different levels of significance and effect sizes. As seen, the indicators based on temporal variability of amplitude appeared superior in those respects, and thus $\Delta CA$ and ${DP}_{max}$ were selected for subsequent analysis. Individual data points corresponding to each LUS dataset are plotted in the $\Delta CA$ and ${DP}_{max}$ space in Fig.4b. The points corresponding to the datasets without B-lines are clustered in the lower left corner of the plot. Thus, it was hypothesized that a combinatorial indicator of the form $K_1\Delta CA+K_1{DP}_{max}$, rather than an individual indicator would have the highest predictive value for presence of B-lines.

Figure 4.

(a) Analysis of RF signal metrics based on cumulative amplitude at large depth from wide and focused beam acquisitions (${CA}_W$ and ${CA}_F$), and its temporal variability at the long (seconds) time scale - $\Delta {CA}_F$ and signal variability over short (milliseconds) time scale - $DP$. The metrics are presented as mean and standard deviation calculated over all LUS datasets. The metrics with the largest effect size — ${DP}_{max}$ and $\Delta {CA}_F$ — were selected for further analysis of predictive value and are shown in (b); each point corresponds to a LUS dataset. (c) Mean cross-validated AUC curve (solid red line) 0.91 (95% CI = 0.80, 0.94) and k-fold ROC curves (dashed lines) quantifying the predictive capabilities for a combinatorial indicator $K_1\Delta CA+K_2{DP}_{max}$.

Predictive Model Derivation

Logistic regression was used to estimate the combined ability of $\Delta {CA}_F$ and ${DP}_{max}$ to predict the presence or absence of B-lines, as determined by expert consensus. Overall cross-validated AUC was 0.91 (95% CI = 0.80-0.94) indicating excellent discrimination between true positives and true negatives (Figure 4c).²³ In order to calculate diagnostic test characteristics, a predictive probability threshold of 0.32 was selected by clinicians to optimize for high sensitivity to impaired pulmonary function, which is clinically useful in an emergency department and hospital setting. A subset of performance metrics for different predictive probability thresholds is shown in Table 2. At this threshold, the associated sensitivity was 93.2%, specificity was 58.5%, the positive predictive value (PPV) was 0.80, and the negative predictive value (NPV) was 0.83. The associated rates of true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN) can be found in the contingency table (Table 3). Note that depending on the clinical context, a different threshold for classifying B-lines as positive or negative may be used.

Table 2.

Performance of Signal Indicator-Based B-line Classification Model at Different Thresholds

Predictive probability	Sensitivity	Specificity
0.1	100	7.3
0.2	98.6	41.5
0.3	94.5	58.5
0.7	69.9	87.8
0.8	64.4	87.8
0.9	47.3	90.2

Table 3.

Classification Accuracy of Signal Indicator-Based B-line Model (Threshold = 0.32)

Predicted	Actual
	No	Yes
No	TN: 24 (21.1%)	FN: 5 (4.4%)
Yes	FP: 17 (14.9%)	TP: 68 (59.7%)
Total (N)	29	85

Discussion

We investigated the feasibility of detecting B-lines in LUS from indicators computed from RF signals rather than reconstructed images. One recent study suggests that semi-automated measurement of B-line artifacts is feasible in ED patients, and may provide important prognostic information for patient care.²⁴ The present study was motivated by the clinical need for a simple wearable, non-imaging LUSS that could enable continuous quantitative feedback to providers and patients on the presence of lung water that indicates acute alveolar or interstitial abnormality. This approach would remove the system and operator dependence inherent to LUS and could be useful in a number of clinical scenarios. Those include treatment dosing for cardiogenic pulmonary edema patients, continuous monitoring of lung condition in pneumonia while protecting providers from contagion, facilitating fast semi-automated triage of patients in emergency settings, titration of ventilator settings in ARDS, determination of ventilator weaning strategies and ongoing monitoring for ventilator induced complications.^{3,8,11-12, 25-26}

In the present study, seven unique candidate RF signal features associated with pulmonary B-line formation were identified from a pilot dataset of LUS videos and corresponding RF data collected in participants with confirmed pulmonary edema and healthy volunteers. The analysis used the RF signals summed across the elements of the LUS probe to mimic a single-element sensor with a size similar to the footprint of a standard phased array. Two indicators associated with RF signal amplitude variability over the long and short time scales were found to be most indicative of B-lines, reflecting the transient shimmering quality of B-lines and their relative motion with respiration. A combination of the two indicators was analyzed for predictive performance, and the overall cross-validated AUC was 0.91, corresponding to excellent discrimination between true positives and true negatives. The computation of the combinatorial indicator is simple, and could be readily automated when used with a single-element LUSS. Combined with the recently reported feasibility of robust and durable acoustic coupling of ultrasound sensors to the skin via adhesive gel pads,²⁷ these findings motivate further development of LUSS.

There were a few notable limitations to this study. Although 160 individual LUS video clips were recorded, there was a small number of participants and corresponding LUS datasets from which candidate indicators were derived. Second, we only analyzed the presence and absence of B-lines, and not their number, brightness and quality (e.g. focal vs diffuse) that could be correlated to the severity of the condition. Our study included patients who were admitted to the hospital, and it is possible that the candidate metrics may not perform as well in less sick individuals due to spectrum bias.²⁸ The specific attributes of B-lines are known to be dependent on the clinical context and differ between pulmonary edema, pneumonia, and ARDS. Thus, the candidate signal indicators may need to be reassessed and validated separately in patients with lung abnormalities of etiology other than cardiogenic pulmonary edema explored here. For our predictive analysis, specific values of the coefficients K1 and K2 were not reported to avoid implications that concrete values have been identified. Instead, we focused on reporting the general ability of the combination of $\Delta {CA}_F$ and ${DP}_{max}$ to associate with presence of B lines over varying coefficients as best fit the subset of data being tested during each round of cross-validation, resulting in the estimatation confidence intervals. More generalizable coefficients will be determined with larger sample sizes, and more varied patient populations, in future validation studies. Further, known history interstitial lung disease and lung fibrosis - conditions that are known to also result in formation of B-lines on LUS – were not part of the exclusion criteria in the study design. Due to the differences in the mechanisms of the formation and characteristics of B-lines between edema and fibrosis, the latter could have confounded the results. Fortunately, per retrospective chart review, none of this study participants had a history of fibrosis, however this should be included in the original study design in the future. Lastly, the mechanism for one of the indicators of B-lines - rapid variability of RF signal amplitude at the pleural line – was only hypothesized to be related to respiratory lung sliding, although it appears consistent with reported clinical use of US M-mode to characterize sliding.²⁰ To rigorously confirm this hypothesis, controlled benchtop experiments with lung-mimicking phantoms are warranted.²⁹

In conclusion, automated indicators based on RF signals from a single-element ultrasound sensor can facilitate the detection of lung water similarly to how B-lines are interpreted in LUS imaging currently, yet bypass the need for multi-element probes, image reconstruction and interpretation.