Machine Learning-Enabled Gas Sensor Based on MOF-Derived In2O3-CuO for Exhaled CO Detection

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: Not formally reported as an effect size; best-performing single sensor (3In2O3-CuO) demonstrated highest response magnitude to CO relative to acetone, with response defined as Ra/Rg (air resistance / gas resistance ratio)

• Primary Outcome: Qualitative discrimination between CO and acetone (C3H6O) using a five-sensor In2O3-CuO array with LDA machine learning; cross-validation accuracy of 82% for CO identification across 1–50 ppm range

• Nnt Or Sensitivity: Clinical sensitivity and specificity not calculable from this study; analytical LOD = 1 ppm CO; response time = 5 seconds; recovery time = 21 seconds; cross-validation classification accuracy = 82%

• Confidence Interval: Not reported for any outcome measure

Clinical Application

Clinical feasibility is premature to assess. Significant engineering, regulatory, and clinical validation work is required before bedside deployment. Challenges include: (1) collecting reproducible breath samples from neonates; (2) managing the highly humid and variable matrix of real exhaled breath; (3) achieving >95% diagnostic accuracy required for screening; (4) device miniaturisation, sterilisation, and single-use or reusable design; (5) integration with clinical workflows in neonatal intensive care or postnatal wards. In Australia, neonatal jaundice is managed according to RANZCOG and RACP/RACP-endorsed guidelines, with serum total bilirubin and transcutaneous bilirubinometry (TcB) as the established diagnostic standards. Exhaled CO breath testing is not currently listed on the Medicare Benefits Schedule (MBS) or recommended in RACGP or RACP clinical guidelines for neonatal jaundice. The TGA would require Class IIb or Class III medical device registration for a diagnostic breath analyser intended for neonatal use, necessitating substantial clinical evidence well beyond this proof-of-concept study. No PBS or MBS pathway exists for this technology at present. Australian neonatal units, particularly in rural and remote settings, could theoretically benefit from a noninvasive, low-cost screening tool if clinical validation is achieved, given the logistical challenges of laboratory bilirubin testing in those environments. Theoretically applicable to neonates with suspected hyperbilirubinaemia where noninvasive CO breath testing could supplement or replace invasive bilirubin sampling; also potentially relevant to adult CO monitoring in occupational or environmental health contexts. Currently applicable only as a research tool.

Abstract

Noninvasive, precise breath analysis holds significant importance for the screening and monitoring of neonatal jaundice, with its core challenge lying in the specific detection of the key biomarker carbon monoxide (CO) within complex breath matrices. This research employed a controlled process to fabricate five sets of In2O3-CuO sensors derived from bimetallic metal-organic frameworks (Bi MOFs). These were assembled into a microsensor array, and by incorporating a linear discriminant analysis algorithm, an electronic nose system was constructed. This approach utilizes thermodynamic feature engineering to enhance data validity and optimize algorithm selection, thereby reducing reliance on large-scale data and computational resources. By integrating thermodynamically feature-driven machine learning, the electronic nose system-comprising merely five In2O3-CuO sensors-ultimately achieved accurate discrimination between CO and acetone. Concurrently, the single 3In2O3-CuO sensor exhibits excellent reproducibility, moisture resistance, and stability, with a low detection limit of 1 ppm and relatively rapid response/recovery times (5/21 s). By introducing machine learning algorithms to analyze the multidimensional response signals from the sensor array, the study successfully addressed the cross-sensitivity issue between CO and coexisting interfering gas C3H6O in clinical environments, achieving qualitative identification of CO ranging from 1 to 50 ppm (cross-validation accuracy reached 82%), providing a highly reliable technical platform for noninvasive screening of neonatal bilirubin metabolic abnormalities.

References

1. 1.Zhao, F., Jiang, Z., Jiang, X., Wang, R., Yu, L., & Wang, H. (2026). Machine learning-enabled gas sensor based on MOF-derived In2O3-CuO for exhaled CO detection. ACS Sensors. Advance online publication. https://doi.org/10.1021/acssensors.6c00346