amr-mic-changepoint: Change-point detection on MIC percentile time series for carbapenem resistance early warning

Background: Antimicrobial resistance (AMR) is a top-ten public health threat. Current surveillance relies on binary minimum inhibitory concentration (MIC) breakpoint classifications, which miss sub-threshold distributional shifts that may presage emerging resistance by 1–3 years. We evaluated whether change point detection on upper-percentile MIC time series provides an objective early warning signal for carbapenem resistance emergence.

Methods: We retrospectively analysed Escherichia coli and Klebsiella pneumoniae tested against imipenem and meropenem in Israel and South Africa (2012–2023), using 10,258 isolates from the Vivli AMR Register across eight pathogen–drug–country combinations. For each stratum-year we extracted seven log₂(MIC) distributional features (P75–P97 percentiles, isolate-weighted geometric mean, IQR). Change points were detected with PELT (RBF kernel) using a per-series BIC-tuned penalty. Performance was assessed by F1, precision and recall against pre-specified CLSI emergence events across predictive (1–3 year lead), concurrent, and combined (0–3 year) windows. Stability used 500 year-resampled bootstraps; Mann–Kendall tested monotonic trends.

Results: Two of eight combinations yielded interpretable signals, both in K. pneumoniae from South Africa. For imipenem, four log₂(MIC) features (P85, P90, P97, geometric mean) tied at F1 = 0.80 in the predictive window (precision = 1.00, recall = 0.67); the P90 series captured all three emergence events in the combined window (F1 = 1.00, mean lead 1.8 years), with the 2017 change point recovered in 74% of bootstraps. For meropenem, the log₂(MIC) IQR predictively flagged all three events (F1 = 1.00, mean lead 1.3 years). Mann–Kendall confirmed significant increasing trends (p < 0.001) only in these two combinations.

Conclusions: All four E. coli combinations and both K. pneumoniae–Israel combinations showed no emergence events, consistent with stable resistance or Israel’s national CRE containment programme. PELT change point detection on log₂(MIC) features from open surveillance data can generate verifiable early warning signals for carbapenem resistance in high-burden settings, via a reproducible open-source pipeline.

Keywords: antimicrobial resistance; change point detection; carbapenem resistance; Klebsiella pneumoniae; early warning; AMR surveillance; South Africa; Israel

If you use this software, please cite both the article and the software.