Ceftriaxone (CRO) is a first-line antibiotic for pediatric infections caused by nontyphoidal Salmonella (NTS). The rising prevalence of ceftriaxone-resistant NTS strains poses a growing public health concern because resistance can compromise empiric and directed therapy in children. Conventional antimicrobial susceptibility testing (AST) provides gold-standard resistance determinations but is inherently time-consuming, producing delays in actionable results for clinicians.
MALDI-TOF MS has been widely adopted in clinical microbiology for rapid and reliable microbial identification using protein spectral fingerprints. However, standard MALDI-TOF MS workflows do not inherently provide antimicrobial susceptibility predictions. The study described here explored whether routinely generated MALDI-TOF MS spectra could be combined with machine learning to provide a rapid, preliminary prediction of ceftriaxone resistance in NTS, potentially shortening the interval to a resistance-informed clinical decision.
The analysis used isolates collected from three medical centers. In total, the study included 472 NTS isolates that yielded 1,414 MALDI-TOF MS spectra. These spectra constitute the input features for model training and validation. The abstract does not report further breakdowns of isolate sources, serotypes, patient demographics, or the precise preprocessing steps applied to raw spectra; those details were not reported in the abstract.
The investigators evaluated a machine learning pipeline using the Light Gradient Boosting Machine (LGB) algorithm to predict CRO resistance from MALDI-TOF MS spectral features. A compact model using only seven features was developed to balance predictive performance and model simplicity. The choice of LGB and the selection of seven features were central to the reported model development; additional modeling choices, feature engineering methods, and hyperparameter settings were not detailed in the abstract and would require consulting the full text for specification.
Performance was assessed across training, internal validation, and two external validation datasets. Reported area under the receiver operating characteristic curve (AUC) values for the seven-feature LGB model were:
These results indicate strong discriminatory ability in training and internal validation with attenuation of AUC on external datasets, consistent with the challenges of model generalizability across centers and sample sets. The abstract does not provide sensitivity, specificity, predictive values, calibration metrics, decision thresholds, or confusion matrices; those details were not reported in the abstract.
To illustrate potential laboratory use, the authors packaged the trained model into an interactive Streamlit web interface. The web interface demonstrates proof-of-concept accessibility and could be used to facilitate further evaluation in laboratory settings. The abstract does not describe integration with laboratory information systems, runtime performance, user workflows, or regulatory considerations; such operational details were not reported in the abstract.
The authors framed this work as a proof-of-concept demonstrating that routinely generated MALDI-TOF MS spectra can be leveraged with machine learning (LGB) to provide rapid preliminary predictions of ceftriaxone resistance in NTS. Key limitations noted in the abstract include the need for further large-scale, prospective, multicenter validation before the approach could be implemented as a routine clinical susceptibility testing tool. The abstract does not report prospective clinical impact, cost–benefit analysis, or post-deployment performance in routine diagnostic workflows; these aspects require future study.
A Light Gradient Boosting Machine model trained on MALDI-TOF MS spectra from 472 isolates (1,414 spectra) produced promising AUCs for predicting ceftriaxone-resistant nontyphoidal Salmonella, with a compact seven-feature model achieving AUCs of 0.94 (training), 0.88 (internal validation), 0.79 and 0.80 (external validations). The model was demonstrated via a Streamlit web interface as a usability proof-of-concept. The study supports further prospective, multicenter evaluation to determine generalizability, diagnostic accuracy metrics in routine practice, and clinical utility prior to adoption as a diagnostic adjunct.