Interpretation of cardiovascular responses to pharmacologic agents in preclinical studies is largely driven by manual review of systolic blood pressure (SBP), pulse pressure, and heart rate (HR) data from radio-telemetry, making blinded assessment of novel compounds subjective, slow, and difficult to standardize across studies. We developed a machine learning framework to learn pharmacologic response signatures from telemetry data and provide an alternative to expert-consensus interpretation.
The dataset comprised 28 dosing sessions from 8 rats implanted with arterial blood pressure radio-telemetry units, covering four mechanistically distinct compounds: phenylephrine, prazosin, ivabradine, and dexamethasone. Sessions were aligned to dosing time, and 68 features were extracted to capture response magnitude, kinetics, temporal shape, and cross-signal interactions. Model development used leave-one-animal-out cross-validation to evaluate generalization to unseen animals while preventing information leakage.
In the matched-control setting, the framework achieved 100% classification accuracy across drug classes. In the unmatched deployment setting, with no paired vehicle, baseline, or cage-mate data, classification from raw aligned telemetry also achieved 100% leave-one-animal-out accuracy, demonstrating that robust pharmacologic discrimination is possible even when controls required for interpretation are unavailable. A confidence-based rejection mechanism distinguished drug sessions from placebo and no-dosing recordings, supporting its use as a screening tool rather than a forced classifier.
In a blinded compound evaluation, the framework identified a mixed pharmacologic profile, with HR appearing more ivabradine-like and SBP appearing more prazosin-like, illustrating how unknown compounds can be interpreted through signal-level and overall-profile similarity rather than a single label. After unblinding, the compound was confirmed as a beta blocker, and the inferred profile matched the expected pharmacologic effect, supporting the framework's ability to recover properties of unknown drugs.
These results suggest that machine learning can transform telemetry analysis from a consensus-based interpretive task into a reproducible, quantitative, and scalable approach for mechanistic screening of unknown compounds.