The researchers developed a framework called tsMIST, or Time Series Model Sensitivity Test, to examine how predictions change along a controlled path between real examples belonging to different classes. Rather than relying only on overall accuracy or aggregate measures of robustness, tsMIST focuses on the point at which a classifier switches its prediction from the class associated with one signal to the class associated with another. This transition is treated as an estimate of the model’s decision boundary along the chosen morphing path. The approach is intended to reveal how much of a transformation a model can tolerate before its interpretation changes, and whether that tolerance corresponds to meaningful changes in the underlying temporal signal.

Time series are sequences in which the order and timing of observations carry information. A medical waveform, for example, may contain a characteristic peak, pause or oscillation whose position is as important as its numerical amplitude. Two signals can have similar values but differ in when their important features occur, while signals with different amplitudes may share the same temporal pattern. This creates a problem for conventional sensitivity analysis. If an artificial test example is generated by simply averaging the values of two signals at each time point, it may change amplitude while destroying the timing relationship that makes the original signals meaningful. A model can then appear sensitive or insensitive for reasons caused by the test itself rather than by its genuine behavior.

The first transformation examined in the study was linear interpolation. Given two time series, linear morphing gradually blends their corresponding measurements. If one signal is represented by a sequence of values from time point one to time point n and another signal has the same structure, a parameter moving from zero to one controls the mixture: at one end, the result is entirely the first signal; at the other, it is entirely the second. Intermediate values produce weighted averages. This method is simple, fast and widely understandable, but it assumes that corresponding positions in the two sequences represent comparable events. That assumption fails when a discriminative feature has shifted in time, when one waveform is stretched relative to another or when the class distinction is defined by temporal alignment rather than raw amplitude.

To address that limitation, the researchers introduced a path-interpolation operator called PathI. Instead of blending the amplitudes at fixed positions, PathI uses correspondences identified through dynamic time warping, a technique designed to align patterns that unfold at different speeds or at slightly different times. Dynamic time warping searches for an alignment between two sequences by allowing one portion of a signal to correspond to a longer or shorter portion of another, while preserving the order of events. PathI then moves matched features through time along that alignment. The resulting intermediate examples are intended to retain more of the temporal structure of the source signals, making the path between classes less like a numerical average and more like a gradual temporal transformation.

The distinction became clear in experiments designed around a known decision boundary. The analysis showed that linear morphing is structurally incapable of recovering the true boundary when the difference between classes is temporal rather than amplitude-based. In such cases, the midpoint produced by averaging may not represent a meaningful intermediate state at all. PathI, by contrast, recovered the boundary exactly under the conditions examined by the authors. The researchers also formally established when the average transition point calculated by tsMIST estimates the classifier’s boundary position. Their results challenge a common interpretation of 0.5 as a universal benchmark: a midpoint value reflects symmetry between the source and target classes, not an inherently correct or ideal location for every classifier or dataset.

The framework summarizes each set of morphing experiments using two measures, tsMISTAvg and tsMISTStd. The average records where prediction switches tend to occur along the transformation path, providing an estimate of the decision boundary’s position. The standard deviation describes how consistently those switches occur across tested pairs or repetitions. A model that changes class near the same point repeatedly may have a stable boundary along the tested direction, whereas a broad spread of transition points indicates less consistent behavior. Importantly, the two measures answer different questions. A classifier may have a boundary positioned close to the midpoint but still behave inconsistently, or it may switch consistently at a location far from the midpoint. Treating these properties as interchangeable could obscure important weaknesses.

The researchers then tested the approach on ten real-world medical time-series datasets and four classifiers. Across those evaluations, InceptionTime consistently produced the most robust decision boundaries, while the Catch22 feature-based approach was the most fragile. The comparison between PathI and linear morphing revealed systematic differences in estimated boundary position, confirming that the operator used to construct synthetic examples affects the apparent sensitivity of a model. At the same time, the researchers found no practically meaningful difference in consistency between the two operators as measured by their variability. PathI nevertheless more closely preserved the temporal structure of the original signals, suggesting that its estimates may be easier to interpret when timing carries the central biological or physical meaning.

The findings do not show that one classifier is universally safe or that a model with a favorable tsMIST score will automatically perform reliably in clinical practice. The experiments assess behavior along selected paths between borderline real instances, and the meaning of robustness depends on which transformations are plausible in the application. A temporal shift that represents natural variation in one medical signal could indicate a serious measurement error in another. Likewise, a morphing path that is mathematically smooth may not correspond to any physiological process. The study therefore frames sensitivity analysis as a structured evaluation problem: researchers must justify how intermediate examples are generated, report the operator used and distinguish changes caused by the model from artifacts introduced by the test procedure.

That requirement could become increasingly important as machine-learning systems move from laboratory benchmarks into settings where small temporal changes have consequences. A classifier that recognizes an arrhythmia, detects a motor disorder or monitors an industrial process may encounter signals that differ from training examples not because their class has changed, but because events occur earlier, later, faster or slower. Testing only with point-by-point averages could miss these failure modes. By comparing linear and correspondence-aware transformations, tsMIST offers a way to ask a more precise question: does the model change its decision because the signal’s meaningful structure has changed, or because the evaluation method has distorted that structure? The authors conclude that morphing operators should never be treated as neutral technical details. They are part of the definition of the robustness experiment itself, and boundary-aware sensitivity analysis should be considered a component of responsible artificial-intelligence evaluation.