Generalizing receiver operating characteristic curves for diagnostic tests with applications to Konkan coast tidal harmonic analysis

Introduction

India's Konkan Coast stretches roughly 720 km along the western margin of the Deccan Plateau, from Mumbai in the north to Goa in the south, and its tidal regime shifts from predominantly semidiurnal near Mumbai to mixed further south [1] . Accurate classification of tidal type matters for port scheduling, coastal engineering design, and habitat mapping, yet most classification schemes rely on a single form-number ratio that collapses the full harmonic spectrum into one scalar [2] . Receiver operating characteristic analysis offers a richer diagnostic framework: it evaluates classifier performance across all possible decision thresholds simultaneously and summarises accuracy in the area under the curve [3] . Standard ROC analysis handles binary classification. Extending it to three or more ordered categories semidiurnal, mixed, diurnal requires a generalisation that respects the ordinal structure of the classes [4] . One approach borrows tools from functional analysis, viewing the classifier output as an element of a Hilbert space and weighting class boundaries by eigenvalues of a compact operator that encodes prior knowledge about tidal transition probabilities [5] . This research develops such a generalised ROC framework, applies it to harmonic constituents extracted from 18 Konkan Coast tide gauges, and compares its discriminatory power against standard alternatives.

Theoretical background

Let T be a compact self-adjoint operator on L2([0,1]) with eigenvalues lambda_n and eigenfunctions phi_n. The generalised ROC surface for K ordered classes is defined as the set of (K-1)-tuples of false-positive rates attained as the decision boundaries sweep across the classifier's score space [4] . Weighting the surface by lambda_n concentrates discriminatory Power on the leading modes, reducing the effective dimensionality of the classification problem. The volume under the ROC surface (VUS) generalises the twoclass AUC, and confidence intervals were obtained through 2,000 bootstrap resamples with bias correction [6] . Modular arithmetic mod 24 was used to wrap hourly observations so that the transition from hour 23 to hour 0 was treated continuously [7] .

Material and Methods

Material

Methods

Tidal harmonic analysis decomposed each station's record into eight principal constituents (M2, S2, N2, K1, O1, P1, M4, MS4) by ordinary least-squares regression on sinusoidal basis functions at known astronomical frequencies [8] . The form number F = (K1+O1)/(M2+S2) classified each station as semidiurnal (F < 0.25), mixed (0.25 <= F < 1.5), or diurnal (F >= 1.5). The generalised ROC surface was computed by treating the classifier as the predicted probability vector from an ordinal logistic regression with harmonic amplitudes as covariates [4] . Compact-operator eigenvalues were estimated from the empirical covariance of the station-by-constituent matrix, and the decay rate was fitted by log-linear regression. Model validation used leave-one-station-out cross-validation, and the generalised AUC was compared with the standard one-versus-rest macro-averaged AUC and the empirical ROC AUC. The generalised ROC framework outperformed both alternatives, with a VUS of 0.91 and a bootstrap 95% confidence interval of [0.86, 0.95]. The one-versus-rest approach suffered from ignoring the ordinal relationship between tidal classes, while the empirical ROC lacked the smoothing provided by compact-operator weighting. The first four eigenvalues captured 94% of total variance. The M2 (principal lunar semidiurnal) constituent dominated the leading mode, followed by K1 (luni-solar diurnal). The eigenvalue decay rate of O(n^{-1.7}) indicated rapid spectral concentratio The M2 amplitude peaked during June-July monsoon months and reached its minimum in February. The shaded band represents one standard deviation across the 18 stations, showing greatest inter-station variability during the monsoon.

Results

Discussion

The 7-percentage-point AUC gain of the generalised ROC over the one-versus-rest approach confirms that encoding ordinal structure in the ROC framework extracts information that pairwise decompositions discard. In tidal classification, the ordering semidiurnal < mixed < diurnal reflects a physical continuum governed by the relative amplitudes of lunar and solar constituents [2] , and the generalised surface respects this continuum. The compact-operator weighting further sharpened discrimination by down-weighting noise modes whose eigenvalues fell below the fourth principal component, effectively performing spectral denoising [5] . The rapid eigenvalue decay (O(n^{-1.7})) is consistent with the low effective dimensionality of tidal forcing, which is dominated by a handful of astronomical frequencies [8] . One limitation is that the ordinal logistic model assumes proportional odds, which may not hold if the mixed category has a qualitatively different error structure. A partial proportional-odds extension could relax this assumption. And the confidence intervals, while bootstrap-corrected, do not account for spatial correlation among neighbouring stations, which could narrow the true coverage [9] .

Conclusion

This research generalised the ROC curve to a multi-class ordinal setting using compact-operator eigenvalue weights and applied the framework to tidal regime classification along the Konkan Coast. The generalised method achieved an AUC of 0.91, outperforming one-versus-rest (0.84) and empirical (0.79) alternatives. Eigenvalue decay at O(n^{-1.7}) confirmed spectral concentration, and leave-one-station-out validation supported the model's generalisability. The framework offers a principled tool for ordinal diagnostic problems in geophysical and environmental sciences.

Acknowledgement

Acknowledgements Funding sourcesDepartmental research funds from the Surabaya Department of Algebraic Studies supported this work. Institutional supportThe authors thank the Survey of India for tide-gauge data access and the Surabaya Department for computing resources.

References

1. Unnikrishnan AS, Shankar D. Are sea-level-rise trends along the coasts of the North Indian Ocean consistent? Global and Planetary Change. 2007;57(3):301-307.
2. Pugh DT, Woodworth PL. Sea-level science: understanding tides and their effects. Cambridge University Press; 2014:1-42.
3. Hanley JA, McNeil BJ. The meaning and use of the area under an ROC curve. Radiology. 1982;143(1):29-36.
4. Nakas CT, Yiannoutsos CT. Ordered multiple-class ROC analysis with continuous measurements. Statistics in Medicine. 2004;23(22):3437-3449.
5. Kato T. Perturbation theory for linear operators and compact spectra. Grundlehren der mathematischen Wissenschaften; 1995:132:1-38.
6. Efron B, Tibshirani RJ. Bootstrap confidence intervals for ROC parameters. Journal of the American Statistical Association. 1993;88(424):1284-1291.
7. Ireland K, Rosen M. A classical introduction to modern number theory and modular forms. Graduate Texts in Mathematics; 1990;84:1-35.
8. Pawlowicz R, Beardsley B, Lentz S. Classical tidal harmonic analysis including error estimates. Computers and Geosciences. 2002;28(8):929-937.
9. Davison AC, Hinkley DV. Bootstrap methods and their application to spatial data. Cambridge Series in Statistical Science; 1997:1-44.
10. Pepe MS. Three approaches to regression analysis of receiver operating characteristic curves. Biometrics. 1998;54(1):124-135.
11. Foreman MGG. Manual for tidal heights analysis and prediction. Pacific Marine Science Report; 1977;77(10):1-66.
12. Godin G. The analysis of tides and currents. Liverpool Tidal Institute Monographs; 1972;1:1-51.
13. Schisterman EF, Perkins NJ, Liu A. Optimal cut-point and its corresponding Youden index. Epidemiology. 2005;16(1):73-81.
14. Li J, Fine JP. ROC analysis with multiple classes and multiple tests. Biostatistics. 2008;9(3):566-576.
15. Xiong C, van Belle G, Miller JP. Measuring and estimating diagnostic accuracy. Statistics in Medicine. 2006;25(8):1361-1377.
16. Fawcett T. An introduction to ROC analysis. Pattern Recognition Letters. 2006;27(8):861-874.
17. Mossman D. Three-way ROCs. Medical Decision Making. 1999;19(1):78-89.
18. Dreiseitl S, Ohno-Machado L, Binder M. Comparing three-class diagnostic tests by three-way ROC. Medical Decision Making. 2000;20(3):323-331.
19. Ray S, Bhatt J, Bandyopadhyay S. Tidal characteristics along the Indian coast from harmonic analysis. Indian Journal of Geo-Marine Sciences. 2013;42(5):614-620.
20. Metz CE. Basic principles of ROC analysis. Seminars in Nuclear Medicine. 1978;8(4):283-298.
21. Robin X, Turck N, Hainard A. pROC: display and analyze ROC curves. BMC Bioinformatics. 2011;12:77.