Paired calibration of the Lume 1.2 sensor against two EPA-approved reference methods — Colilert (IDEXX defined-substrate, MPN) and membrane filtration (MF, CFU) — and the Aquagenx Compartment Bag Test (CBT, MPN) used in international monitoring, for E. coli and total coliform quantification.
Field Validation
Lab Validation
The customer dashboards (Boulder, Denver, DC, future deployments) all load a single shared model module at /js/ecoli-model.js, served from shared/ecoli-model.js in the virridy-lume-summary repo. The two models below are what those dashboards actually evaluate against incoming sensor readings; everything else on this page is calibration evidence supporting them.
Continuous prediction
mon2_corrected = mon2 · exp(−ρ·(T−20)) · exp(−k·NTU)log₁₀(cfu) = b₀ + FE[barcode] + β·mon2_corrected
Fit on post-burn-in field data (n=24, 4 sensors). ρ = −0.0792, k = +0.00193. NTU derived from signal_per_spad_kcps. Sensor FE applied for calibration sensors; new deployments fall back to the reference intercept.
ECOLI_CFU_COEFS
Binary alert ≥126 CFU/100 mL
P(≥126) ~ mon2 + floor_temp + tof_mean. Lab-fit logistic regression on raw features (no correction baked in). Threshold selected by Youden’s J on LOO-CV probabilities.
ECOLI_COEFS
Live data from the mWater Lume 1.2 – 2026 Validation Data datagrid. Each water sample collection event is paired with a reference enumeration; the Method column distinguishes which reference was used — Colilert (IDEXX defined-substrate MPN), membrane filtration (MF, CFU), or compartment bag tests (CBT, MPN). The Use in Calibration column flags rows that are unusable because no /diagnostics record (water temperature, required by the CFU regression) was streaming within ±20 min of the sample. All date/time columns are displayed in UTC. Times are corrected from the mWater-stored value using the Timezone Entered column: mWater records times using the data-entry device’s local clock (Boulder, MDT = UTC−6); for samples collected in a different timezone the stored time is adjusted accordingly.
All unique Colilert grab samples collected at Boulder Creek sites (n = 39, deduplicated). Values in CFU/100 mL. EPA single-sample recreational threshold: 126 CFU/100 mL.
| Site | n | Min | Median | Max | ≥126 CFU | All values (CFU/100 mL) |
|---|---|---|---|---|---|---|
| BC-CU | 12 | 12 | 47 | 1986 | 2 (17%) | 12, 15, 15, 21, 26, 44, 50, 53, 60, 75, 152, 1986 |
| BC-55 | 13 | 6 | 53 | 866 | 4 (31%) | 6, 20, 23, 24, 26, 28, 53, 75, 131, 145, 166, 378, 866 |
| BC-30 | 3 | 36 | 105 | 517 | 1 (33%) | 36, 105, 517 |
| BC-Can | 8 | 2 | 16 | 30 | 0 (0%) | 2, 3, 5, 7, 16, 17, 27, 30 |
| BC-Eben | 3 | 6 | 28 | 30 | 0 (0%) | 6, 28, 30 |
| All BC | 39 | 2 | 30 | 1986 | 7 (18%) | median = 30 • mean = 171 • ≥126: 7 of 39 |
Sensor signal is run through a physics-motivated correction pipeline derived from Bedell et al. 2022 (temperature) and Skinner et al. 2024 (turbidity), with ρ and k fit empirically from this field dataset:
mon2_corrected = mon2_val · exp(−ρ · (sipm_temperature − 20)) · exp(−k · NTU), NTU = max(0, −145.89 + 2.0488 · signal_per_spad_kcps)
Single-predictor OLS with per-sensor fixed-effect intercept: log10(colilert) ~ barcode + mon2_corrected. Source data: ⬇ field_matched_512.csv
3-feature logistic regression (the same structure used by the production binary classifier and the lab section below). Features: mon2_val, sipm_temperature, tof_mean (z-scored per LOO fold) plus per-sensor fixed-effect dummies for non-reference field sensors. LOO-CV. Threshold = Youden-optimal.
Jan 2–6, 2026 calibration sessions (paper training range). Fluorescence signal (mon2_val) at the paper operating point: led_power = 1024, sipm_bias = 3040. CBT calibration data is in the field-calibration table above.
⬇ Download full dataset (CSV)
This is the binary alert classifier deployed to all production dashboards (Boulder, Denver, DC). Logistic regression on the full lab dataset (Jan 2–21 2026, n = 300), binary label Colilert ≥ 126 CFU/100 mL. Features: mon2_val_512, floor_temp, tof_mean (z-scored). Performance estimated by leave-one-out cross-validation. Single-event caveat: the 9 positive examples all come from the Jan 8 contamination event (three consecutive 25-min time slots), so when any one positive row is held out the other 8 from the same event remain in training. Positive-class sensitivity is over-optimistic for an unseen new event.
Pooled OLS (Jan 2–6, n = 125): log₁₀(colilert) ~ barcode + signal × floor_temp × tof_mean. Barcode is a fixed-effect intercept shift (reference: 50030); slopes shared. Fit separately for each LED/bias operating point. In-sample R².
LED = 1024 · bias = 3040 — paper
LED = 512 · bias = 3000 — production
LED = 256 · bias = 3300 — original
How does the Lume compare against the two EPA-approved laboratory methods — Colilert (IDEXX) and membrane filtration (MF) — and against itself when retrained on a different reference? Each column analyzes paired samples across three frameworks: log-log regression (top), Bland-Altman agreement (middle), and categorical classification (bottom).
The dedicated method comparison study pairs Colilert (n = 2 replicates) with membrane filtration (n = 3 replicates) across 161 datetimes; 8 zero-valued pairs are excluded from the log-scale analysis, yielding 153 observations. The two EPA-approved methods show R² = 0.572 with a +0.35 log10 bias — MF systematically reads ~2.2× higher than Colilert. 95% limits of agreement span [−0.64, +1.34], meaning paired lab samples can differ by up to ~22× in either direction. Categorical accuracy is 0.66 (Cohen’s κ = 0.40), i.e. “fair” agreement. This inter-method disagreement sets the ceiling for what any sensor can be expected to achieve against either reference.
The Colilert-trained Lume regression is evaluated against Colilert across all bench (n = 176) and field (n = 33) observations. The sensor achieves R² = 0.881, a bias of 0.00 log10, and tight limits of agreement [−0.42, +0.42] — Lume predictions stay within ~2.6× of the reference. Categorical accuracy is 0.89 with κ = 0.88, which is “almost perfect” agreement. Against its training reference, the Lume performs as well as or better than the two EPA methods perform against each other.
The same Colilert-trained Lume model is now evaluated against membrane filtration — a reference method it was never trained on. Performance drops to R² = 0.514 with LoA [−0.80, +0.83] and categorical accuracy 0.84 (κ = 0.65). Critically, the ~0.37 drop in R² from column 2 to column 3 is of the same order as the inter-method disagreement between Colilert and MF themselves (column 1, R² = 0.572). Most of the apparent loss is attributable to reference-method disagreement, not sensor limitations.
To isolate the effect of reference-method choice, the Lume regression is refit using MF as the training target, over the full bucket dataset. Performance jumps back to R² = 0.872 — essentially matching the Colilert-trained model against Colilert. Bias is 0.00 with LoA [−0.93, +0.93]; the slightly wider LoA reflects the higher within-method variability of MF replicates (57.9% RPD vs. 43.5% for Colilert), not a sensor deficiency. Categorical accuracy is 0.81 (κ = 0.66).
Sensor-to-reference agreement is bounded by reference-method reproducibility, not by Lume hardware. Whichever culture method is adopted as truth, the Lume fits it at R² ≈ 0.87–0.88. The gap between columns 2 and 3 is almost exactly the disagreement between the two lab methods themselves (column 1). The Lume is method-agnostic; its ceiling is set by the reference it is trained against, and it already achieves quantitative performance at or above the inter-method agreement ceiling between the two accepted laboratory techniques — while providing continuous temporal coverage that grab-sample laboratory methods cannot.