Artificial Intelligence: A Bayesian
Probability Principle-Based Algorithm
for Bladder Cancer
Authors:
Dale Srinivas*, Sarika Gupta (NHS, UK), Anu Singh, Arvind Kumar Gupta (NHS, UK), Rakesh
Kumar Garg, Shivanand K. Geer, and Terry Feroze
Affiliations:
Twin Island University, Bushlot, Guyana (Dale Srinivas, Anu Singh, Rakesh Kumar Garg,
Shivanand K. Geer, Terry Feroze)
NHS, UK (Sarika Gupta, Arvind Kumar Gupta)
Abstract
We extend our Bayesian algorithm for bladder cancer by providing explicit mathematical
formulations for risk priors, likelihood updating, stage inference, recurrence modelling,
expected-utility treatment selection, surveillance optimisation, and value-of-information
analysis. Worked numeric examples illustrate practical implementation.
1. Mathematical Framework
1.1 Priors for Prevalent Cancer and Stage
Let Ca denote prevalent bladder cancer and MIBC denote stage ≥T2 given cancer. Logistic
priors:
logit P0(Ca) = β0 + β1·(VisibleHaematuria) + β2·Age + β3·PackYears + β4·OccupExposure +
…
logit P0(MIBC | Ca) = α0 + α1·(SolidLesion) + α2·(Size>3cm) + α3·(Hydronephrosis) +
α4·(VariantHistology) + …
1.2 Likelihood Updating with Diagnostic Evidence
For evidence E1..Ek with likelihood ratios LR(Ei), assuming conditional independence:
PosteriorOdds(Ca | E) = PriorOdds(Ca) × ∏ LR(Ei)
P(Ca | E) = PosteriorOdds / (1 + PosteriorOdds)
When tests are dependent, use a Bayesian network to specify P(E | Ca) jointly or apply a
dependence correction factor ρ (0<ρ≤1):
Adjusted LR ≈ [∏ LR(Ei)]^ρ
1.3 Stage Updating After Pathology/Imaging
PosteriorOdds(MIBC | Ca, Estage) = PriorOdds(MIBC | Ca) × ∏ LR_estage(Fj)
Examples Fj: MRI/CT signs of extravesical extension, LVI, high-grade T1 with CIS, size >3
cm.
1.4 Recurrence and Progression Modelling
For NMIBC, let R(t) denote recurrence by time t. Two pragmatic models:
(A) Exponential time-to-event with rate λ: S(t)=exp(−λt). Place Γ(a,b) prior on λ ⇒
posterior Γ(a+events, b+time).
(B) Cox model: h(t|X)=h0(t)·exp(γᵗX). Place priors on γ and estimate via partial likelihood
or Bayesian MCMC.
Convert EORTC/CUETO scores to priors for λ or γᵗX; update with early post-treatment
markers (e.g., 3-month cystoscopy, cytology/FISH) using LR markers.
1.5 Expected-Utility Treatment Selection
For options o ∈ {BCG, EarlyCystectomy, Trimodality}, define utilities U over outcomes
(QALY-weighted):
EU(o) = Σ_outcomes P(outcome | evidence, o) · U(outcome)
Where outcomes include: cure, recurrence, progression to MIBC, treatment complications,
treatment mortality.
Choose argmax_o EU(o).
1.6 Pauker–Kassirer Decision Thresholds
Let B_t be expected benefit of treating true disease, H_t the harm of treating when disease
absent.
Treatment threshold pt = H_t / (B_t + H_t). If P(disease) ≥ pt, treat without further testing.
Test threshold pT satisfies: Expected value(no test) = Expected value(test).
1.7 Surveillance Interval Optimisation
Let c_CYS be cost/disutility per cystoscopy and c_MISS cost of a missed early
recurrence/progression. For interval Δ:
Expected Loss(Δ) = c_CYS · (T/Δ) + c_MISS · P(recurrence between visits | evidence)
Choose Δ that minimises Expected Loss subject to clinical constraints (e.g., Δ ∈ {3,6,12
months}).
1.8 Value of Information (VOI)
EVPI = E[ max_o U(o) | perfect information ] − max_o E[ U(o) | current information ]
EVSI(Test) ≈ E[ max_o E[U(o) | current + test] ] − max_o E[U(o) | current ] − Cost(Test)
Order tests by highest EVSI per cost/time.
1.9 Calibration and Discrimination
Assess calibration via Brier score and calibration slope/intercept; assess discrimination via
AUC/ROC. Apply isotonic or Platt scaling to correct miscalibration.
2. Worked Numeric Examples
2.1 Sequential Updating Example (Revisited)
Prior P0(Ca)=0.18 ⇒ Odds0=0.2195. Cytology LR+=10 ⇒ Odds1=2.195 ⇒ P1=0.687.
Cystoscopy LR+=15 ⇒ Odds2=32.925 ⇒ P2=0.970.
2.2 Expected Utility for High-Risk NMIBC: BCG vs Early Cystectomy
Assume posterior P(occult MIBC)=0.35. Utilities (QALYs over horizon):
U(Cure after cystectomy)=1.00, U(Cure after BCG)=0.95, U(Progression on BCG)=0.60,
U(Complication severe)=0.70, U(Death)=0.00.
Event probabilities:
BCG: P(cure)=0.55, P(recurrence no progression)=0.25, P(progression)=0.20; P(severe
complication)=0.05.
Early cystectomy: P(cure)=0.75, P(complication severe)=0.10, perioperative
mortality=0.02.
Illustrative EU(BCG) ≈ 0.853; EU(Early Cystectomy) ≈ 0.820. Prefer option with higher EU.
2.3 Surveillance Interval Choice
Let c_CYS=0.01 QALY-equivalent per cystoscopy, c_MISS=0.20 QALY loss if recurrence
occurs between visits.
For Δ=3 months over T=12 months ⇒ 4 cystoscopies ⇒ cost=4×0.01=0.04. If P(recurrence
within any 3-month interval)=0.08 ⇒ miss loss=0.016.
Total loss=0.056. For Δ=6 months: 2 cystoscopies ⇒ cost=0.02; P(recurrence in 6
months)=0.14 ⇒ miss loss=0.028; total=0.048 ⇒ prefer Δ=6 months.
2.4 EVSI Example: Cytology vs FISH
Suppose current decision is indeterminate (two options have similar EU). Expected
improvement in EU after cytology is 0.015 QALY, Cost=0.002 ⇒ EVSI=0.013. For FISH:
improvement 0.018, Cost=0.010 ⇒ EVSI=0.008. ⇒ Choose cytology first.
3. Figures
Figure 1. Bayesian Updating Curve
Figure 2. Recurrence Risk Over Time
Figure 3. Treatment Algorithm Flowchart
4. Tables
Table 1. Likelihood Ratios of Common Investigations
Test LR+ (Positive) LR- (Negative)
Urine Cytology (HG) 12.0 0.65
UroVysion FISH 8.0 0.60
CT Urogram (mass) 10.0 0.50
Cystoscopy (lesion) 15.0 0.20
Table 2. Bayesian Utility-Based Decision for High-Risk NMIBC
Posterior P(MIBC) Recommended Treatment Expected Utility
<20% BCG maintenance High QoL, lower morbidity
20–50% Shared decision: BCG vs Balanced
early cystectomy
>50% Radical cystectomy Best oncological control
References
Babjuk, M., Burger, M., Compérat, E.M. et al. (2022) ‘EAU Guidelines on Non–Muscle-
Invasive Bladder Cancer (TaT1 and CIS)’, European Urology, 81(1), pp. 75–94.
Alfred Witjes, J., Bruins, H.M., Cathomas, R. et al. (2022) ‘EAU Guidelines on Muscle-invasive
and Metastatic Bladder Cancer’, European Urology, 82(1), pp. 1–25.
Sylvester, R.J., van der Meijden, A.P.M., Oosterlinck, W. et al. (2006) ‘Predicting recurrence
and progression in individual patients with stage Ta T1 bladder cancer using EORTC risk
tables’, European Urology, 49(3), pp. 466–477.
Fernández, M.I., López-Beltrán, A., Vázquez-Alonso, F. et al. (2018) ‘New molecular
biomarkers in bladder cancer’, Histology and Histopathology, 33(4), pp. 385–400.
Lotan, Y. and Roehrborn, C.G. (2003) ‘Sensitivity and Specificity of Commonly Available
Bladder Tumor Markers Versus Cytology: Results of a Comprehensive Literature Review
and Meta-Analyses’, Urology, 61(1), pp. 109–118.
Pearl, J. (1988) Probabilistic Reasoning in Intelligent Systems: Networks of Plausible
Inference. San Francisco: Morgan Kaufmann.
Cowell, R.G., Dawid, A.P., Lauritzen, S.L. and Spiegelhalter, D.J. (1999) Probabilistic Networks
and Expert Systems. New York: Springer.
Esteva, A., Robicquet, A., Ramsundar, B. et al. (2019) ‘A guide to deep learning in healthcare’,
Nature Medicine, 25(1), pp. 24–29.
Burnham, K.P. and Anderson, D.R. (2002) Model Selection and Multimodel Inference: A
Practical Information-Theoretic Approach. New York: Springer.
Kattan, M.W., Zlotta, A., Shariat, S. et al. (2004) ‘External validation of a Bayesian network
predictive model for bladder cancer’, BJU International, 93(6), pp. 764–769.
Appendix: Symbols and Definitions
Symbol Definition
Ca Presence of bladder cancer (any stage).
MIBC Muscle-invasive bladder cancer (stage
≥T2).
NMIBC Non–muscle-invasive bladder cancer (Ta,
T1, CIS).
P0 Prior probability before new evidence.
Odds Defined as P / (1−P), used for LR updating.
LR+ Likelihood ratio of a positive test =
Sensitivity / (1−Specificity).
LR− Likelihood ratio of a negative test =
(1−Sensitivity) / Specificity.
Posterior Probability Updated probability after incorporating
evidence.
logit(p) Logit transformation: ln(p / (1−p)).
EU(o) Expected utility of treatment option o,
based on outcome probabilities and utility
weights.
QALY Quality-adjusted life year, a measure of
health utility.
EVPI Expected Value of Perfect Information.
EVSI Expected Value of Sample Information (a
test).
λ Hazard rate in exponential survival model.
S(t) Survival probability at time t.
h(t|X) Hazard function at time t for covariates X
(Cox model).
Δ Surveillance interval (months).
c_CYS Cost/disutility per cystoscopy.
c_MISS Cost of missing a recurrence/progression
between visits.