Spirometry Forced Vital Capacity

Iit Gandhinagar, India

Abstract—Respiratory illnesses are a significant global health burden. Respiratory illnesses, primarily Chronic ob- structive pulmonary disease (COPD), is the seventh leading cause of poor health worldwide and the third leading cause of death worldwide, causing 3.23 million deaths in 2019, necessitating early identification and diagnosis for effective mitigation. Among the diagnostic tools employed, spirome- try plays a crucial role in detecting respiratory abnormal- ities. However, conventional clinical spirometry methods often entail considerable costs and practical limitations like the need for specialized equipment, trained personnel, and a dedicated clinical setting, making them less accessible. To address these challenges, wearable spirometry technologies have emerged as promising alternatives, offering accurate, cost-effective, and convenient solutions. The development of machine learning models for wearable spirometry heavily relies on the availability of high-quality ground truth spirometry data, which is a laborious and expensive en- deavor. In this research, we propose using active learning, a sub-field of machine learning, to mitigate the challenges associated with data collection and labeling. By strategically selecting samples from the ground truth spirometer, we can mitigate the need for resource-intensive data collection.

We present evidence that models trained on small subsets obtained through active learning achieve comparable/better results than models trained on the complete dataset.

Index Terms—Machine Learning In Healthcare, Ac-

tive Learning, Cost-effective Respiratory Diagnostic Tools,

I. Introduction

Respiratory illnesses pose a significant health challenge worldwide. Every year around 262 million people suffer from asthma, and 3.23 million deaths are attributed

To Chronic Obstructive Pulmonary Disease (Copd)

With the emergence of a global pandemic like COVID- 19, monitoring lung health has become an essential part of human survival . Early detection and diagnosis of respiratory illnesses can mitigate their impact and reduce the risk , .

Studies have indicated that the diagnosis of lung-related conditions is often missed or delayed until the health conditions have significantly worsened and clear signs of poor lung health are evident. One contributing factor to this issue is the relatively high cost of medical tests, which typically costs $2401. Consequently, the availability of affordable and user-friendly equipment for diagnosing lung health is limited and poses a significant obstacle to timely diagnosis. .

Spirometry is a non-invasive pulmonary function test widely used to diagnose respiratory problems, and assess lung function . Typically conducted in a clinical set- ting by trained clinician, the test requires certain prepara-

Tory Measures Such As Avoiding Smoking, Consuming

heavy meals, or engaging in tedious physical activity, as these factors can affect the accuracy of the results . During the test, the patient exhales forcefully into a spirometer. The spirometer measures 23 different lung function parameters, such as forced vital capacity (FVC), peak expiratory flow (PEF) and forced expiratory volume in one second (FEV1), providing insights into respiratory health.

Although spirometry helps assess the severity of lung diseases, performing clinical spirometry is expensive.

Mobile And Wearable Spirometry , , Pose

as affordable alternatives to clinical spirometry. These are cost-effective and have higher portability, enhancing accessibility. However, these approaches also have their drawbacks as accuracy is affected by factors like posi- tioning of the device, the user’s lip posture, and external

Environmental Conditions ,

Verifying these techniques requires ground truth spirom- etry data, which is difficult to obtain due to logistical challenges in collecting consistent and reliable mea- surements across diverse environments and is hugely impacted by participants’ cooperation and adherence to standardized testing procedures. Thus datasets used to test out mobile spirometry approaches like SpiroS-

Mart And Mobspiro Are Limited In Participant

size. The same trend is observed across majority of the 1Source: https://www.vernier.com/product/spirometer/

14

Body sounds like eating, breathing, speech etc.

Fecg Data

TABLE I: Comparison of dataset size across various physiological signal based studies. The general trend across the table showcases the unavailability of large size datasets in the domain. physiological datasets. As seen in table I, most prior studies related to physiological signals have been work- ing with small amounts of data. As per , collecting a spirometry sample takes about 45 minutes to an hour and relies heavily on the patient’s ability to follow clinical instructions, which can be challenging, especially for young and elderly patients.

The core premise of active learning is that a machine learning system can improve its accuracy if it is allowed to select the data it learns from. In this research, we propose an active learning approach using a publicly available spirometry dataset , which contains esti-

Mated Values For Fev1, Fvc, And Pef Obtained From

a wearable spirometry study. We aim to investigate the benefits of active learning by selecting the most infor- mative samples that lead to a reduction in errors from previous baselines. To evaluate our approach, we explore different learning strategies like Single Task and Multi-

Task Active Learning And Single Output And Multi-

Output (which are further discussed in the upcoming sections). The American Thoracic Society (ATS) allows up to 7% error rate between spirometry maneuvers .

Our proposed methods consistently outperform this with error rates of 4.96% for FEV1 and 4.45% for FVC, using Leave-One-Out cross-validation. We achieve these errors using active learning with ∼30% of the total dataset in the individual task learning setting (single-task setting).

Since collecting participant data specific to individual tasks [FVC, FEV1] is impractical, we propose leveraging the joint distribution across output parameters to select

Common Points Across Both Parameters. By Applying

various weighting schemes and utilizing models that inherently support multi-output settings (such as Ran- dom Forest and Decision Trees), we can strategically leverage the correlation between outputs and improve performance.

To summarize, the main contributions of this research

Work Are:

• Active Learning for Spirometry: We propose using ac- tive learning for spirometry to reduce data acquisition costs while ensuring same or better performance using a subset of data (∼24 participants).

Schemes That Can Use The Correlation Between The

output parameters and jointly optimize over them to reduce errors across tasks. • Reproducibility: We publicly release our codebase and believe that our work is fully reproducible. All the generated tables and graphs have corresponding scripts

Ii. Background

This section discusses about the pertinent background for clinical spirometry, the recent advances in mobile and wearable spirometry and briefly introduces the concept of active learning.

A. Clinical Spirometry

Spirometry is a pulmonary test widely used to assess breathing patterns and detect lung ailments like asthma, cystic fibrosis, and COPD . During a spirometry ma- neuver a patient breathes into a spirometer mouthpiece which measures the amount and speed of air exhaled.

Spirometers processes the collected flow-volume curves (Figure 1) to extract lung paramters. The American Thoracic Society (ATS) discusses different lung health 2Source: https://github.com/percom-anonymous/Spiroactive

Pef

Fig. 1: Graphical Representation of the flow-volume curve in spirometry showing the three commonly used lung health indices : Peak Expiratory Flow (PEF), Forced Expiratory Volume in 1 second (FEV1) and Forced Vital Capacity (FVC).

parameters that are extracted from spirometry. The three

Majorly Used Parameters Are:

• FVC - Forced vital capacity is the total air volume exhaled (Litres).The plot of the expired volume illus- trates the FVC as the peak of the graph, representing the maximum expiration volume.

• FEV1 - Forced expiratory volume in 1 second is the exhaled air volume in the first second of exhalation (Litres).

• PEF - Peak expiratory flow is the maximum airflow velocity in exhalation. (L/s). The plot of the flow rate illustrates this at the peak value in the expiration stage.

B. Mobile And Wearable Spirometry

Previous efforts have investigated alternative systems

, Due To Challenges In Clinical Spirometry Such

as: i) Difficulty in ensuring patient cooperation, es- pecially with young children and elderly patients; ii) Spirometry instrumentation is expensive and requires time-consuming efforts for data collection. To address these issues, cheaper and more accessible alternatives such as wearable and smartphone spirometry have been

Studied In The Past. , , , . Recent Efforts

have also been made to incorporate wearables (such as masks) with audio and pressure sensors to extract vital

Lung Parameters , –. They Provide A More

secure and enclosed environment to capture audio data, that is used as a proxy for the airflow, to estimate lung

Health Indices Such As Fev1, Pef And Fvc. Despite

promising results, studies in this area usually consists of small participant groups (typically <70 as shown in Table I) owing to difficulty in collecting ground truth spirometry data. The efficiency of active learning in low data settings coupled with the convenience of hand held

Update Pool

Fig. 2: Active Learning Loop that showcases the flow of

Data Across The Whole Pipeline

spirometry motivated us towards studying active learning for mobile spirometry.

C. Active Learning

One of the main reasons behind the success of machine learning in the recent past has been the availability of large datasets. Collecting large quantities of unlabeled data in medical and imaging settings is often easy.

However, obtaining the labels for such datasets is usually time-consuming and expensive. Active Learning is a sub-field of machine learning that aims to reduce the labeling or annotation cost by strate- gically labeling samples from the large unlabeled set.

The active learning scenarios can be broadly divided

Into Three Types : I) Membership Query Synthesis;

ii) Stream-Based Selective Sampling and iii) Pool-Based

Sampling. We Direct The Reader To A Comprehensive

active learning survey and focus on the setting most pertinent for our application: pool-based sampling. Pool-based sampling is used in many real-world prob- lems like text classification , information extrac-

Tion , Speech Recognition , And Cancer Diagno-

sis . Pool-based sampling has also been widely used to reduce human annotations in the activity recognition

Task. , Broadly, Pool-Based Sampling Is Based

on labeling samples from a large pool of unlabeled data instances via heuristics called acquisition functions. We now present the active learning workflow for pool-based sampling, as illustrated in Figure 2.

1) The active learning loop consists of a machine learn- ing model initially trained on the labeled dataset (usually small initially).

2) The trained model is then used to make predictions on the large pool set.

3

3) We use acquisition function, A(x), to calculate scores for the pool set. In this context, the variable x denotes the data points within the pool set.

4) We choose the next point to be queried as the one

Argmaxx∗∈Xpool A(X∗)

5) We obtain the true label/ground truth (y∗) for the

Chosen Pool Point ¯X∗

6) We add the point (¯x∗, y∗) to the training dataset and correspondingly remove ¯x∗from the pool dataset. 7) Continue with Step 1 iteratively until the allocated budget, representing the maximum number of queries or labeled samples, is depleted, or until the predefined stopping condition is met within the active learning framework.

For x∗, we can look at uncertainty in output y∗predicted by the current model and use the variance as the acqui- sition function: A (x∗) = var (y∗). Generally, two kinds of methods are used in the active learning literature for

• Uncertainty Sampling: Here We Can Use Machine

learning models that inherently provide uncertainty estimates and use this uncertainty as our acquisition.

Committee Of Multiple Machine Learning Models And

use the variance in prediction across this committee as our acquisition.

Iii. Problem Statement

Mobile/wearable spirometry systems collect audio data while users perform forced breathing maneuvers. The collected raw audio data undergoes normalization and Hilbert transformation to generate its envelope. Various features are extracted from the audio as per and the ground truth lung health indices are collected using a

Forced Breathing Spirometer Maneuver

Collecting wearable audio data is significantly easier as compared to acquiring the ground truth spirometer data, which requires expert supervision and clinic visit. Thus,

We Pose Our Active Learning Problem As Follows:

• We assume a large number of samples (pool data) of audio data and corresponding features collected from mobile/wearable spirometry setups.

• We assume only a small fraction of samples (training data) to be associated with ground truth data contain- ing the two lung health indices.

• Our goal is to “strategically” choose participants from the pool set and obtain their true labels. • The role of the oracle in our active learning setting is akin to visiting the lab where the ground truth spirometry can be performed.

The overall vision is to get “good” performance using a small number of points sampled from the pool. Prior literature considers a 7% mean absolute percentage error between our estimated lung health indices and ground truth to be a “good” performance per the ATS

A. Learning Paradigms

We now present our method, focusing on the primary lung health indices: FEV1, FVC. Prior mobile/wearable spirometry literature has proposed distinct “tuned” fea- tures for each health index based on domain knowledge

. For Example, Fev1 Is The Volume Of Air Exhaled

in the first second of a forced expiratory maneuver, so signals beyond 1s are not needed. In contrast, measuring FVC requires the complete signal. Given the different set of (partially overlapping) features, we consider esti- mating FEV1, and FVC, as two different tasks. For our experimentation, we use the Single Output setting where we train a different model for each individual task. Each trained model, though possibly of the same class has its own distinct parameters (θ) and hyperparameters (ϕ).

Previous literature , has generally focused on this experimental setting.

B. Active Learning Settings

As discussed in section II-C, the active learning loop queries points that improve model performance. The

Querying For Pool Points Can Be Done In Two Ways:

1) Single-Task Active Learning where different points are queried for different tasks. 2) Multi-Task Active Learning where the same point

In This Setting, The Two Tasks (Fev1 & Fvc) Are

treated as two independent problems. The querying for each task, at each iteration, is done based on the given

¯X∗= Argmaxx∗∈Xpool A(X∗)

where A (x∗) = var (y∗) This leads to difference in the order in which pool points are selected in each task. Single Task AL can be achieved by using the single output models as described in section IV-A .

Multi-Task Active Learning

While Single-Task AL tends to achieve the best results as the problems are independent of each other, it is often wasteful to acquire separate participant data for

4

each task. This section discusses the importance of using a single approach to pick points that lead to error reduction across all the tasks. The MultiTask AL acquistion function using weighing can be stated as:

∗) ∗Wm

where M is the number of outputs (e.g. M = 2 for our

∗Is The

output for task m and wm is the weight associated with the mthoutput. We use various weighing techniques to

Determine The Weights Wm:

1) Uniform weighted uncertainty: We uniformly weigh the uncertainties across all the tasks and pick the most uncertain data point from the pool.

M

2) Maximum Weighted uncertainty: The weighing factor

Gives The Highest Weightage To The Task With The

highest uncertainty.

I∗)

where N is the total number of samples in the pool,

I∗Is The

prediction for the mth task corresponding to the xi∗ point from the pool. 3) Inversely weighted uncertainty: The task with the maximum uncertainty is given minimum weightage.

!−1

The other Multi-task AL acquisition functions include: • Rank Sampling: Each data point is ranked by uncer- tainty across models. The ranks are summed, and the point with the highest uncertainty/lowest rank sum is selected for active learning.



• Round Robin Sampling: In this technique, we query the most uncertain points across each task in a round robin fashion, i.e., in the first iteration, the point is queried according to the first task and so on. The round-robin sampling technique ensures that each task is given an equal number of samples, promoting fair- ness and reducing bias in the AL loop.

A. Dataset

For our experimentation, we have collaborated with

Spiromask Authors And Collected 10 More Samples

to expand our dataset. Out of the total 58 samples (including 10 new sample), we remove 7 samples based on the Flow-Volume envelope and the back-extrapolated volume (BEV) .

Out of 51 participants (Table II), 10 suffer from various lung ailments (asthma, obstructive lung ailments, wheez- ing, COVID infections, and other restrictive ailments).

The study consists of 16 female participants, and the age group of people varies from 21 to 68 years . To generate the feature set for active learning, we pass the audio signal through a Butterworth bandpass filter

With Cut-Off Frequencies Of 3000 And 5000 Hz That

conserve all the tidal breathing information. The filtered audio is used to generate acoustic and temporal features like Spectrogram features (STFT),

Mel Spectrogram, Mel-Frequency Energy (Mfe), Mel-

frequency cepstral coefficients (MFCC), and spectral features. We then apply Sequential Forward Selection to jointly optimize our dataset for both the output vari-

, , . Although Pef Is Commonly Used

for lung health prediction, discussions with the dataset authors revealed that a small number of samples were clamped partly impacting the signal and peak required for calculating PEF. However, it does not affect the features needed for FVC and FEV1. Hence, PEF is not considered for future experimentation.

B. Metrics

We use the mean absolute percentage error to compare performance across all the participants for FVC, and FEV1 tasks.

∗100

where y is the true value, ˆy is the predicted value and N is the total number of test points. The guidelines

Describe The Error Between Two Spirometry Efforts

in percentages. Hence we choose MAPE scores as the measurement method.

C. Experiments

In this section, we discuss the experiments and their results to study the efficacy of active learning for spirom- etry. For our experiments, we use a committee consisting of Random Forest Regressor, Decision Tree Regressor,

Related Journal Articles & DOI Links

Selected peer-reviewed publications relevant to 12 Lead ECG Acquisition. Click the DOI to access the full paper (may require institutional access).

Why Choose Us?

Bangalore guidance for robotics, Spectre and autonomous systems projects.