Definition
Health Score Contextual Accuracy (HSCA) measures whether the health scores your APM system produces actually predict proximity to functional failure when you hold operating context constant.
The Measurement Gap
Bus 4471 showed a health score of 78. Bus 4472 showed 71. Both were in service on a Tuesday morning in a mid-sized transit fleet of 1,200 vehicles. The APM system flagged anything below 60 for mandatory inspection: both buses cleared that threshold. Both appeared on the fleet health dashboard as acceptable. Neither was pulled from service.
Bus 4471 broke down at 7:42 a.m. on Route 17. Complete air brake system failure. The route serves a school corridor, and the bus was carrying passengers during peak hours. Nobody was hurt, but the near-miss triggered a full fleet review.
Here is what the post-incident analysis found: Bus 4471 operated exclusively on Route 17, a hill-heavy urban route with 31% more braking events per kilometer than the fleet average. The health scoring algorithm that generated that 78 was trained on fleet-wide data. It didn't account for route-specific duty cycle. A score of 78 on Route 17 represents a fundamentally different brake condition than a score of 78 on a flat suburban route. The algorithm treated every bus as if it ran the same route. It didn't.
Bus 4472, with its lower score of 71, ran a flat suburban route with gentle grades and moderate braking frequency. Retrospectively, its actual brake condition was better than Bus 4471's: despite the lower score. The number lied because the context was invisible.
The fleet engineer pulled historical data and ran the comparison. For flat routes, a health score in the 75–80 band correlated with an average of 47 days before brake intervention. For hilly routes, that same 75–80 band correlated with 12 days. Same number. Wildly different operational meaning. The fleet had been treating both as equivalent because the dashboard showed both in green.
This is not a fringe scenario. It is the default failure mode of health scoring systems that aggregate condition data without stratifying by operating context. Load, environment, duty cycle, and asset vintage all affect how a condition indicator maps to actual proximity to failure. A pump running at 110% of rated capacity degrades differently than one at 70% capacity: even when their vibration signatures score identically. A compressor in a high-humidity coastal environment accumulates fouling at rates that a score trained on inland climate data will underestimate. An asset that's been in service 18 years responds to condition changes differently than a four-year-old asset of the same model.
Health score dashboards show green because they are doing exactly what they were designed to do: aggregate condition indicators into a single number and compare that number against a threshold. What they are not doing (and what they cannot do without context stratification) is tell you whether that number means the same thing for this asset in this operating environment as it does for the fleet average.
Asset Availability captures breakdowns. OEE captures lost production. Neither tells you that the breakdown traced back to a health score that was structurally incapable of reflecting actual condition for that asset class in that operating context. The dashboard said 87% of the fleet was above threshold on the morning Bus 4471 failed. That statement was technically accurate and operationally misleading at the same time.
There is no metric that tells you whether an automated health score accurately reflects proximity to functional failure given specific operating context.
That gap is what Health Score Contextual Accuracy is designed to fill. It doesn't improve your health scoring algorithm directly. It tells you which contexts your algorithm handles well and which ones it fails in, so you know when to trust the score and when to treat it with skepticism.
Formula
Health Score Contextual Accuracy (HSCA) measures whether the health scores your APM system produces actually predict proximity to functional failure when you hold operating context constant. It answers the question Bus 4471's fleet manager should have been asking every morning: does an 85% health score mean the same thing for this asset running this duty cycle as it does for the fleet average?
The Formula
HSCA = 1 − (σ_RUL|H,C / μ_RUL|H,C)
Where:
- σ_RUL|H,C = standard deviation of actual remaining useful life for assets with health score H in operating context C- μ_RUL|H,C = mean actual remaining useful life for assets with health score H in operating context C- H = health score band (e.g., 75–80%)
- C = operating context category (load level, environment, duty cycle, vintage)
This is 1 minus the coefficient of variation of actual RUL, segmented by health score band and operating context. The logic is direct: if a health score is contextually accurate, then all assets with that score in that context should reach functional failure at roughly similar times. Low variance relative to the mean means the score is predictively consistent. High variance means the same score is covering fundamentally different condition states, and you can't act on it reliably. Higher HSCA is better. A score of 1.0 would mean perfect consistency: every asset at that health score in that context fails at exactly the mean RUL. A score approaching 0 means the health scores are noise relative to actual failure timing in that context. The thresholds: HSCA at or above 0.70 (coefficient of variation at or below 0.30) indicates the score is contextually meaningful and can support maintenance scheduling decisions. Between 0.50 and 0.70, it's directionally useful but not precise enough for tight scheduling. Below 0.50, the score is effectively meaningless without context adjustment: you're better off using a fixed-interval schedule.
Worked Example
The transit fleet, segmented by health score band and route type, produced the following results after 14 months of historical analysis:For the 75–80 health score band on flat routes: mean RUL of 47 days, standard deviation of 11 days, coefficient of variation of 0.234, HSCA of 0.766. That's contextually meaningful, a bus scoring 75–80 on a flat route will, with reasonable consistency, need brake intervention within seven weeks. For the 75–80 band on hilly routes: mean RUL of 12 days, standard deviation of 9 days, coefficient of variation of 0.750, HSCA of 0.250. Nearly useless. A bus scoring 75–80 on a hilly route could have anywhere from 3 to 21 days of remaining life. You cannot schedule reliably from that number. For the 75–80 band on mixed routes: mean RUL of 28 days, standard deviation of 18 days, coefficient of variation of 0.643, HSCA of 0.357. Directionally useful at best: the score tells you intervention is coming, but not when. For the 60–65 band on flat routes: mean RUL of 18 days, standard deviation of 5 days, coefficient of variation of 0.278, HSCA of 0.722. Still contextually meaningful: the flat route environment produces consistent degradation even at lower health scores. For the 60–65 band on hilly routes: mean RUL of 4 days, standard deviation of 3 days, coefficient of variation of 0.750, HSCA of 0.250. Same problem as the higher band. The hilly route context introduces so much variance in actual degradation rate that the health score can't track it. The pattern that emerges: the health scoring algorithm works reasonably well for the operating context it was implicitly trained on (flat routes, average duty cycle). It fails for the tail contexts (higher duty cycle, steeper grades, more braking events) where the underlying physics diverge from the training population mean.
Edge Cases
Small sample within a context group is the first problem you'll hit. If you have fewer than five assets with a given health score band in a given operating context, the variance estimate is statistically unreliable. Exclude those cells from HSCA calculation and flag them as "insufficient data for context." Don't interpolate, an unreliable HSCA is worse than no HSCA because it creates false confidence. Mean RUL approaching zero is the second. When assets in a context group are already very close to failure, the coefficient of variation becomes erratic: small absolute deviations produce large relative values. For context groups where mean RUL is below approximately five days (or the equivalent threshold for your asset class), switch to absolute deviation rather than CoV. Report the absolute spread in days rather than a ratio. Censored data is the third. Assets that received condition-based interventions before reaching functional failure can contribute to HSCA calculation only if the intervention event revealed actual condition state: inspection records, teardown findings, or diagnostic results that let you estimate what the RUL would have been. Assets pulled for scheduled replacement with no condition assessment cannot contribute. Health scores not comparable across platforms is the fourth. If your organization runs multiple APM systems with different scoring scales, normalize to a common 0–100 scale before segmenting. Mixing a system that scores 0–1 with one that scores 0–100 in the same HSCA calculation will produce meaningless results.
Cross-References
HSCA sits downstream of two KPIs that determine the quality of health score inputs. Digital Twin Fidelity Index (DTFI, GMI-AR-019) measures how accurately your digital twin simulates real asset behavior: health scores derived from twin simulations inherit whatever fidelity problems exist in the twin, and those fidelity gaps tend to be context-specific. Model Drift Detection Rate (MDDR, GMI-AR-018) tracks how quickly your predictive models deteriorate as operating conditions change, a drifting model will produce increasingly context-insensitive health scores, and HSCA will detect the consequence even when model drift metrics aren't being watched. HSCA feeds two downstream KPIs. Operator Trust Calibration Score (OTCS, GMI-AR-021) measures whether maintenance teams trust predictive recommendations at the right level: contextually inaccurate health scores, left uncorrected, erode operator trust faster than almost any other failure mode because the disconnect between the score and the field reality is visible to technicians even when it's invisible to dashboards. Degradation Projection Validity Rate (DPVR, GMI-AR-007) measures what percentage of RUL projections prove accurate within tolerance: any degradation projection built on a contextually inaccurate health score inherits that inaccuracy directly, producing RUL estimates that miss the mark.
Proposed formula, pending validation pilots.
At a Glance
Decision Supported
Does "85% health" actually mean what I think it means for this specific asset in its specific operating context, or is the health score misleading because it doesn't account for context?