PaperlayerView exact revision on arXiv

Source-linked AI summary

AI4COVID-19: AI Enabled Preliminary Diagnosis for COVID-19 from Cough Samples via an App

Ali Imran, Iryna Posokhova, Haneya N. Qureshi, Usama Masood, Muhammad Sajid Riaz, Kamran Ali, Charles N. John, MD Iftikhar Hussain, Muhammad Nabeel

arXiv:2004.01275v6eess.AScs.LGcs.SDq-bio.QMstat.ML

TL;DR

The paper presents a ubiquitously deployable AI-based preliminary COVID-19 diagnosis tool using cough sounds through a mobile app. Its reported results are promising, while the authors identify the need for medically supervised real-world validation and systematic large-scale cough data gathering.

  • Problem

    Isolating infected people is described as the only effective defense against COVID-19, while the lack of it contributes to historic losses.

  • Method

    AI4COVID-19 is a mediator-centered AI engine for cough-based preliminary COVID-19 diagnosis delivered through a mobile app.

  • Results

    The results show that the AI4COVID-19 app can diagnose COVID-19 with negligible misdiagnosis probability.

  • Takeaways & Limitations

    The tool is presented as a functional option for timely, cost-effective, and safe monitoring, tracing, and tracking.

  • Takeaways & Limitations

    The current version requires medically supervised real-world validation, and some data were not included in training and testing.

Abstract

from arXiv · show

Background: The inability to test at scale has become humanity's Achille's heel in the ongoing war against the COVID-19 pandemic. A scalable screening tool would be a game changer. Building on the prior work on cough-based diagnosis of respiratory diseases, we propose, develop and test an Artificial Intelligence (AI)-powered screening solution for COVID-19 infection that is deployable via a smartphone app. The app, named AI4COVID-19 records and sends three 3-second cough sounds to an AI engine running in the cloud, and returns a result within two minutes. Methods: Cough is a symptom of over thirty non-COVID-19 related medical conditions. This makes the diagnosis of a COVID-19 infection by cough alone an extremely challenging multidisciplinary problem. We address this problem by investigating the distinctness of pathomorphological alterations in the respiratory system induced by COVID-19 infection when compared to other respiratory infections. To overcome the COVID-19 cough training data shortage we exploit transfer learning. To reduce the misdiagnosis risk stemming from the complex dimensionality of the problem, we leverage a multi-pronged mediator centered risk-averse AI architecture. Results: Results show AI4COVID-19 can distinguish among COVID-19 coughs and several types of non-COVID-19 coughs. The accuracy is promising enough to encourage a large-scale collection of labeled cough data to gauge the generalization capability of AI4COVID-19. AI4COVID-19 is not a clinical grade testing tool. Instead, it offers a screening tool deployable anytime, anywhere, by anyone. It can also be a clinical decision assistance tool used to channel clinical-testing and treatment to those who need it the most, thereby saving more lives.

2 AI4Lyf LLC, USA

The paper motivates scalable COVID-19 screening because clinical testing is difficult to provide rapidly, affordably, safely, and at population scale. Existing molecular and imaging approaches remain constrained by access, facilities, turnaround time, exposure risks, or diagnostic trade-offs.

  • Motivation: Scalable proactive screening is needed because some infected people do not seek care and can still act as active spreaders.The paper links this need to population-scale pre-screening, self-isolation, and directing medical care toward those needing it most.
  • Testing constraints: Clinical testing is limited by availability, expense, required facility visits, turnaround time, and risks to patients and medical staff.The paper describes geographical and temporal availability constraints, scarce and costly tests, hospital or clinic visits, delays, and infection risks for healthcare workers.
  • Testing constraints: Rapid molecular tests still require office visits and do not solve many access and distancing problems.The paper also notes shortages of specimen-collection equipment and continuing facility requirements despite faster results.
  • Testing constraints: At-home sample collection can introduce shipping delays, sample-quality compromises, and self-collection errors.Samples must be shipped to authorized laboratories, while storage and untrained collection may affect results.
  • Alternative approaches: Imaging-based diagnosis can improve sensitivity but may have low specificity and still requires a well-equipped clinical facility.The cited CT evidence reports pooled sensitivity of 94% and specificity of 37%, while both imaging approaches retain facility-visit requirements.

B. Proposed cough based COVID-19 screening approach

The proposed approach tests whether COVID-19 cough contains distinguishable features and turns that hypothesis into a mobile, cloud-based preliminary screening system. It combines medical analysis, cough data, and risk-averse classification to support accessible screening.

  • Medical rationale: The approach investigates whether COVID-19 causes respiratory alterations and cough features distinct from those of other respiratory diseases.The analysis uses X-rays, CT scans, autopsy studies, and prior literature to assess this theoretical possibility.
  • Data collection: The study gathers spontaneous and requested COVID-19 coughs alongside pertussis, bronchitis, and normal cough samples.Including requested coughs is intended to extend applicability to infected people who are not yet showing cough as a symptom.
  • Cough analysis: Experiments found COVID-19-associated cough features distinguishable from pertussis, bronchitis, and normal cough.Multiple analysis approaches were used, and the paper reports that the hypothesis was supported in these comparisons.
  • System design: AI4COVID-19 records cough through a user-friendly mobile app and sends it wirelessly to a cloud AI engine for preliminary diagnosis.The engine first detects whether the recorded sound is a cough and can operate amid background noise before passing coughs to diagnosis.
  • Risk control: A risk-averse architecture combines three independently designed classifiers through a mediator, returning “Test inconclusive” when they disagree.The design gives each classifier veto power and trades a lower misdiagnosis rate for more inconclusive results.

II. METHODOLOGY

The methodology addresses whether cough can support COVID-19 diagnosis despite cough’s broad medical causes and limited COVID-19 training data. It uses domain knowledge to narrow the hypothesis and reduce the data required for testing.

  • Hypothesis: The hypothesis is that sufficiently unique latent cough features could support AI-based differential diagnosis relative to non-COVID-19 coughs.The paper treats distinct pathomorphological changes as a prerequisite for meaningful cough-based diagnosis and rejects the idea if such differences are absent.
  • Problem formulation: Cough is caused by many COVID-19-unrelated conditions, creating a substantial risk that an AI system will confuse their coughs with COVID-19 cough.The paper identifies cough as a common symptom across bacterial, viral, respiratory, and non-respiratory conditions.
  • Problem formulation: A brute-force solution would require extensive labeled cough data for every relevant condition, which the paper considers impractical during the pandemic.The authors state that collecting such comprehensive data would take too much time for the immediate problem.
  • Domain-aware design: The domain-aware design uses respiratory medical knowledge to analyze COVID-19 pathomorphology, identify irritant locations, and guide feature extraction.This approach is intended to narrow hypothesis testing and minimize the amount of data needed.
  • Validation strategy: The authors therefore evaluate whether COVID-19 cough features are distinct from coughs associated with common respiratory conditions rather than relying only on blind large-data training.They connect this design choice to the need for timely, feasible validation.

2) Distinct pathomorphological alternations in respiratory system caused by COVID-19:

COVID-19 produces respiratory changes that differ from other respiratory infections, motivating investigation of whether these differences appear in cough sounds. Prior work also supports cough-based diagnosis of several respiratory diseases.

  • COVID-19 pneumonia differs from non-COVID-19 pneumonia across multiple CT features, including peripheral distribution, ground-glass opacity, and vascular thickening.Reported differences also include reverse halo sign, pleural findings, effusion, and lymphadenopathy.
  • COVID-19-related pathomorphological alterations are expected to produce idiosyncratic cough sound signatures.The paper links distinct underlying respiratory changes to potentially distinct cough features.
  • AI analyses of X-ray and CT images have differentiated COVID-19 from bacterial and non-COVID-19 viral infections with good accuracy.These findings further suggest that COVID-19 affects the respiratory system distinctly from other respiratory infections.
  • Cough-based diagnosis has previously achieved over 80% sensitivity and specificity for asthma, pneumonia, bronchiolitis, croup, and lower respiratory tract infections.The cited clinical study supports cough as a diagnostic signal across several respiratory conditions.
  • This work proposes, evaluates, and prototypes a smartphone-app AI engine for anytime, anywhere COVID-19 tele-testing and pre-screening.The authors present it as the first work known to them to develop and test such a prototype.

B. Data Description and Practical Viability of the Solution with Available Data

Because comprehensive disease-specific cough data were unavailable, the study uses domain knowledge, selected comparison conditions, and audio-feature processing to assess practical feasibility. The available data show distinct feature patterns across four cough classes, while the limited sample sizes constrain generalizability.

  • Data Description: The study uses domain knowledge to shortlist pertussis, bronchitis, asthma, pneumonia, bronchiolitis, croup, and influenza as likely confusable conditions.The shortlist reflects similar respiratory pathomorphological changes and potentially similar cough signatures.
  • Data Description: The cough detector uses 1838 cough sounds and 3597 non-cough environmental sounds, including data from ESC-50 and the smartphone app.It filters cough from 50 types of common environmental noise before diagnosis.
  • Data Description: The diagnosis dataset contains 96 bronchitis, 130 pertussis, 70 COVID-19, and 247 normal cough samples from different people.The samples were collected from patients with the named conditions and healthy people.
  • Data Description: The available diagnosis samples are small, and the authors state that more data are needed to make the solution more generalizable.New COVID-19 samples were being used to test the trained algorithm as they arrived.
  • Data pre-processing and visualization: Each audio file is an uncompressed 16-bit PCM recording sampled at 44.1 kHz with a fixed 3-second length and converted to the Mel scale.The Mel representation provides higher resolution at lower frequencies, where cough sounds have more energy.
  • Data pre-processing and visualization: t-SNE visualizations show distinct features across cough types, including COVID-19 versus bronchitis and pertussis, supporting practical feasibility.The authors use this observation to proceed with an AI-engine design for app deployment.

C. The AI4COVID-19 AI-Engine

AI4COVID-19 uses a two-stage cloud-based engine for cough detection and COVID-19 diagnosis. The diagnosis stage includes three classifier variants whose outputs are mediated to support cautious decisions.

  • The system has two stages: detecting cough from mixed sounds and diagnosing COVID-19 from cough.
  • The smartphone app provides the user interface, while the trained cough-detection and diagnosis models run on a cloud server.Cloud deployment allows back-end model refinement without updating the app.
  • The diagnosis block combines deep transfer-learning multiclass, classical machine-learning multiclass, and deep transfer-learning binary classifiers.The architecture is organized around a cough detector followed by these three diagnosis models.

1) System architecture:

The system filters input sounds, sends detected coughs to three parallel classifiers, and reports a diagnosis only when they agree. This mediator-centered design is intended to reduce misdiagnosis risk while supporting real-time app use.

  • System architecture: The smartphone app records and forwards sounds to the server, where a cough detector first filters inputs before diagnosis.Poor-quality or non-cough samples are rejected and the user is prompted to re-record.
  • System architecture: Detected coughs are processed by DTL-MC, CML-MC, and DTL-BC in parallel, with their outputs passed to a mediator.
  • System architecture: The app reports a diagnosis only when all three classifiers agree; otherwise, it returns “test inconclusive.”Each classifier effectively has veto power in the mediator-centered design.
  • System architecture: The authors report extremely low opposite-direction COVID-19 prediction errors on the available testing data.This result is attributed to the tri-pronged architecture despite higher misdiagnosis rates for individual classifiers.

3) COVID-19 diagnosis:

COVID-19 diagnosis uses three parallel classifiers that combine deep transfer learning, classical machine learning, and binary classification. The approaches use cough representations including Mel spectrograms, MFCCs, and PCA features, with cross-validation and regularization for limited data.

  • 3) COVID-19 diagnosis:: Three independently designed classifiers diagnose COVID-19 from detected coughs using parallel deep and classical approaches.The system includes DTL-MC, CML-MC, and DTL-BC classifiers.
  • 3) COVID-19 diagnosis:: DTL-MC classifies coughs into COVID-19, pertussis, bronchitis, or normal respiratory status using Mel spectrograms and a CNN.Its architecture is modified to produce four output classes.
  • 3) COVID-19 diagnosis:: Transfer learning reuses cough-detector features so the diagnosis models can be trained with limited disease-specific data.Initial low-level features are retained while later layers are fine-tuned for subtler disease-related features.
  • 3) COVID-19 diagnosis:: CML-MC uses MFCC and PCA feature extraction followed by a multi-class SVM, with class balancing and k-fold validation over 100,000 iterations.This branch uses a different preprocessing path from the spectrogram-based classifier.
  • 3) COVID-19 diagnosis:: The models are evaluated with accuracy, specificity, sensitivity/recall, precision, and F1-score using cross-validation and regularization.Hyperparameters are tuned against cross-validation accuracy to address limited data and over-fitting.

A. Cough detection

The cough detector identifies cough events with 95.60% overall accuracy, while three diagnosis classifiers provide complementary COVID-19 classification performance. The mediator-based architecture is analyzed under an independence assumption and is reported to reduce misdiagnosis at the cost of inconclusive outputs.

  • A. Cough detection: 95.60% overall accuracy was achieved by the cough detector when distinguishing cough events from non-cough events.The result is based on the reported confusion matrix and performance metrics.
  • A. Cough detection: Deep learning-based classifiers outperformed the manual-feature classical machine-learning classifier, with transfer learning and automatic feature extraction cited as contributing factors.The comparison is between DTL-MC and DTL-BC versus CML-MC.
  • C. Overall performance under independence assumption: Under the independence assumption, the mediator combines three classifier decisions to reduce overall false-positive and false-negative probabilities.The architecture uses an inconclusive result to counter over- or under-sensitivity among individual classifiers.
  • C. Overall performance under independence assumption: The app currently predicts an inconclusive result 38.7% of the time, while the analysis notes that the classifiers are not truly independent in reality.Dependence may arise from shared demographic factors and similarities in classifier architectures.

IV. DISCUSSION

AI4COVID-19 is presented as a supplement to clinical testing rather than a replacement, with uses in tele-screening, access-limited settings, and exposure reduction. The app also supports scalable screening, social distancing, and anonymized monitoring applications.

  • IV. DISCUSSION: The app is not intended to replace or compete with medical-grade testing.Its proposed role is complementary preliminary screening and clinical decision assistance.
  • IV. DISCUSSION: AI4COVID-19 can enable tele-screening anywhere and help address shortages of testing facilities, especially in remote areas.The paper describes possible use as a clinical decision assistance tool where medical resources are limited.
  • IV. DISCUSSION: The app may reduce unnecessary medic exposure for non-critical patients advised to stay home or self-isolate.This use is framed as a complement to medical care rather than a substitute for clinical testing.
  • IV. DISCUSSION: Anonymously spatio-temporally tagged cough samples and centralized test records could support monitoring, long-term medical planning, and policy making.The paper also links the app to informed social distancing and self-isolation.
  • IV. DISCUSSION: The system is proposed as a low-cost supplement to temperature screening because it can run on existing smartphones and process non-spontaneous coughs.The paper also identifies possible deployment at airports, borders, and other locations.

B. Comparison and contrast of AI4COVID-19 with existing studies

AI4COVID-19 is positioned as an anywhere, anytime cough-screening approach that avoids the clinical visits required by existing testing and imaging methods. Its promising preliminary performance is constrained by limited and potentially imperfect data, motivating larger medically supervised validation.

  • Existing NAAT methods are more sensitive but face geographical and temporal availability limits, higher cost, longer turnaround, in-person visits, and protective-equipment needs.
  • AI4COVID-19 is presented as usable anywhere, anytime, by anyone, including through commodity smartphones or a browser-accessible web portal.
  • X-ray and CT-based approaches report sensitivities ranging from 72% to 96% but still require visits to well-equipped clinical facilities.
  • Transfer learning and combined deep-learning and manually engineered-feature approaches address limited COVID-19 cough data, but more data remains necessary.
  • The mediator-based architecture declares confusing samples inconclusive near decision boundaries, while future work will add more conditions and patient dimensions.
  • The findings provide preliminary proof of concept, but large-scale, medically supervised real-world trials are needed to evaluate generalization and practical performance.
  • The prototype uses cloud-based processing, with a thin-client smartphone app that records and sends audio data without stringent CPU or RAM requirements.

D. Planned Future upgrades of AI4COVID-19

Planned upgrades aim to improve AI4COVID-19’s accuracy and generalization by adding acoustic, metadata, multisensory, imaging, and wearable-sensor inputs. The paper presents the current system as a proof of concept requiring more labeled data and large-scale trials before population-scale use.

  • Acoustic inputs: Adding breathing sounds and speech to cough recordings may improve AI4COVID-19 accuracy.
  • Metadata: Future evaluation will examine age, gender, smoking status, ethnicity, and medical history as metadata for broader generalization.
  • Multisensory inputs: Multisensory inputs, including skin images and wearable or ambient biomarkers, are proposed to improve diagnostic reliability.Examples include infrared cameras, wireless sensors, wristbands, rings, and skin patches.
  • Scientific basis: The approach builds on differential analysis of COVID-19 respiratory alterations and latent cough features relative to other respiratory conditions.Latent features were examined in coughs associated with COVID-19, pertussis, bronchitis, and non-infectious conditions.
  • Current system: The tri-pronged mediator-centered engine is intended to reduce misdiagnosis risk while distinguishing COVID-19 from non-COVID-19 coughs.
  • Validation and deployment: The system is presented as a proof of concept that needs more labeled data and large-scale trials before population-scale prescreening.The proposed use is timely, cost-effective monitoring and tele-screening, particularly where testing is scarce.
Loading 2004.01275v6…