Stop Guessing: A "Safety-First" Tool for Dalbavancin Triage

• A Support Vector Machine (SVM) classifier trained on ~9,800 simulated pharmacokinetic profiles successfully predicted whether standard two-dose dalbavancin regimens would maintain therapeutic levels through Week 8; the model achieved ~95% accuracy in a small clinical validation cohort (n=38) with zero observed false negatives, outperforming traditional Bayesian forecasting for this specific binary risk assessment.

• The tool utilizes readily available covariates (Age, Weight, CrCL), pathogen MIC, and a single pre-second-dose trough concentration to stratify patients; however, immediate clinical deployment is constrained by the strict requirement for precise (non-censored) MIC values and the logistical lag of send-out dalbavancin assays.

• Off-label use of dalbavancin for deep-seated infections like osteomyelitis and endocarditis requires sustained bactericidal exposure for 6–8 weeks, yet substantial inter-individual pharmacokinetic variability makes dosing durability unpredictable.

• Clinicians currently face a "black box" decision at Week 6: empirically administer a costly third dose (~$3,000) or risk subtherapeutic levels and treatment failure; this uncertainty drives defensive over-prescribing and complicates OPAT discharge planning.

• Resolving this ambiguity requires a validated predictive tool to steward resources and ensure safety, but existing Bayesian methods can be resource-intensive and sensitive to single-point data noise.

• Generated a training dataset of ~9,800 "Digital Twin" profiles using the Carrothers population pharmacokinetic (popPK) model, sampling realistic distributions for Age, Weight, Creatinine Clearance, and MIC.

• Developed and compared multiple machine learning classifiers, ultimately selecting a Support Vector Machine (SVM) due to its superior ability to minimize false negatives (missed underexposures).

• Validated the model using a tiered approach: internal hold-out testing, external stress-testing against simulated profiles from two alternative popPK models (Cojutti, Baiardi), and clinical validation against a retrospective cohort of 38 real-world patients from two French university hospitals.

• Benchmarked performance against the current gold standard, Maximum A Posteriori Bayesian Estimation (MAP-BE), specifically assessing the ability to correctly classify patients as "Therapeutic" or "Subtherapeutic" at Week 8.

• In the real-world clinical validation (n=38), the SVM model achieved an overall accuracy approaching 95% and, most critically, demonstrated 100% sensitivity—identifying every patient who actually dropped below the target threshold (zero false negatives).

• The ML classifier outperformed the Bayesian comparator (MAP-BE) in the binary safety task; in the Limoges cohort, MAP-BE misclassified six underexposed patients as safe (76% accuracy), whereas the ML model flagged all of them correctly.

• Operational trade-offs were evident: to achieve high safety, the model accepted a modest Positive Predictive Value (~60%), implying it will recommend redosing for some patients who might have remained therapeutic, prioritizing relapse prevention over drug savings.

• Simulation data revealed that underexposure prior to Week 5 is statistically negligible, suggesting that monitoring efforts should be strictly focused on the Week 5–8 window.

• The "MIC Trap" is the primary barrier: The model requires precise, continuous MIC inputs (e.g., 0.03 mg/L); inputting standard censored laboratory values (e.g., "<= 0.125") will cause the model to drastically overestimate risk and recommend unnecessary doses for nearly all patients.

• Workflow Latency: Because dalbavancin levels are typically send-out tests with multi-day turnaround times, the Day 8 trough cannot drive real-time dosing changes; the tool functions effectively as a Week 4/Week 6 triage aid rather than a bedside adjustment tool.

• Governance Required: The current implementation relies on a public web interface ("Shiny App"), necessitating internal IT hosting and code validation to meet HIPAA and cybersecurity standards before institutional use.

• Target Selection: The model hard-codes a conservative bactericidal target (fAUC/MIC > 111.1); institutions comfortable with bacteriostatic targets for suppressive therapy may find the tool's redosing recommendations too aggressive.

Strengths:

• Safety-First Architecture: The SVM decision boundary was specifically tuned to eliminate false negatives, aligning perfectly with the clinical imperative to prevent catastrophic treatment failure in deep-seated infections.

• Robust Training Volume: Leveraging ~9,800 simulated profiles allowed the model to learn complex physiological relationships and edge cases that would take decades to accrue in prospective clinical trials.

• Superior Binary Classification: The study highlights that for the specific "Redose: Yes/No" question, a trained discriminative classifier may be more robust to noisy single-point data than generative Bayesian models.

Limitations:

• Validation Size: While the training set was massive, the real-world validation relied on a small sample of 38 patients, limiting certainty regarding performance in diverse populations (e.g., morbid obesity, amputees).

• Single-Model Bias: The ML training data is derived entirely from the Carrothers equation; if a patient's physiology deviates from that underlying mathematical model, the predictions are liable to fail.

• Operational Barriers: The exclusion of albumin as a predictor (for usability) and the strict requirement for E-test/Broth Microdilution MICs render the tool unusable for hospitals relying on standard automated susceptibility reporting.

This research validates machine learning as a highly accurate, safety-biased triage tool for dalbavancin redosing, but implementation is currently stalled by the operational need for precise MIC workflows and internal IT governance.