Cost-Effectiveness, Budget Impact and Value of Information of Artificial-Intelligence-Supported Detection of Breast, Lung, Colorectal and Cervical Cancer in the Philippines: A Decision-Analytic Modelling Study

1. Introduction

The Philippines carries a high and late-presenting cancer burden. The Global Cancer Observatory estimates about 149,852 new cancers and 86,338 cancer deaths in 2024, with breast (28,124 cases), colorectal (17,462), lung (15,312) and cervical cancer (7,829) among the leading sites and lung cancer the leading cause of cancer death [1,2]. In a regional comparison of 2022 estimates, the Philippines recorded the highest breast-cancer mortality rate in south-east Asia and the highest male lung-cancer incidence and mortality [3]. Population-based registry data from Metro Manila and Rizal show that most cancers are diagnosed with regional or distant spread [4,5], and five-year relative survival for Philippine residents in 1998–2002 was 58.6% for breast, 12.0% for lung, 40.2% for colorectal and 45.4% for cervical cancer, far below the survival of Filipino-Americans with the same cancers [6]. Screening participation remains low: in the 2022 National Demographic and Health Survey, 10% of women aged 15–49 reported ever having a provider examination for breast cancer and 9% ever having a cervical screening test [7,8], and a 2023 analysis summarised that about 1% of eligible Filipino women, some 540,000 of 54 million, had been screened for breast or cervical cancer [9]. Treatment capacity is also thin: a 2022 survey counted 53 linear accelerators and 125 radiation oncologists [10], radiotherapy utilisation is about 4.5% of incident cases against an optimal 53% [11], and oncology specialists are concentrated in a few regions [12,13].

Policy has moved. The National Integrated Cancer Control Act (Republic Act No. 11215) mandates organised screening, a population-based cancer registry and expanded PhilHealth benefits [14]. In 2025 PhilHealth introduced outpatient benefit packages for mammography (PHP 2,610), breast ultrasound (PHP 1,350), low-dose chest computed tomography (PHP 7,220) and screening colonoscopy (PHP 23,640), with quality standards for lung and breast screening [15,16,17], and the Health Technology Assessment Council (HTAC) recommended public financing of high-risk HPV DNA testing for women aged 30–65 every five years [18]. A modelled scale-up of a navigated clinical-breast-examination programme has already been found cost-effective, at PHP 60,711 per disability-adjusted life-year averted [19].

Software that interprets screening examinations is a candidate for addressing the reading bottleneck rather than the acquisition bottleneck. Randomised evidence now exists for several relevant tests. In the Swedish MASAI trial, AI-supported mammography screening reduced screen-reading workload by 44%, increased cancer detection by 29% without increasing false positives, and achieved higher sensitivity (80.5% versus 73.8%) with identical specificity (98.5%) and a non-inferior interval-cancer rate [20,21,22]. A pragmatic randomised trial of AI-assisted chest radiography in a health-check population more than doubled the detection of actionable nodules without raising the false-referral rate [23]. Meta-analyses of randomised trials show that computer-aided detection during colonoscopy increases the adenoma detection rate from about 36% to 44–45% and advanced colorectal neoplasia detection by a relative 16%, at the cost of additional non-neoplastic polypectomies [24,25]. AI-assisted cervical cytology raised sensitivity for cervical intraepithelial neoplasia grade 2 or worse (CIN2+) by 5.8 percentage points in a 703,103-woman programme in China [26], and automated visual evaluation of cervical images has been validated prospectively in Africa [27]. In radiotherapy, deep-learning auto-contouring reduces physician delineation time by 30–76% [28,29,30].

Philippine evidence is much thinner. A structured review of applications documented between 2020 and 2026 identified only two retrospective peer-reviewed studies with Philippine patient data, a multinational mammography reader study that included Philippine-affiliated radiologists, commercial chest-radiograph deployments without published local performance, and a cancer-society programme reported only in the press; it found no prospective diagnostic-accuracy or outcome study in Filipino screening participants and no Philippine economic evaluation of interpretive screening software [31,32,33,34,35]. The same review noted that the only Philippine price benchmark, about PHP 50 per AI-read chest radiograph quoted by the Philippine Cancer Society [36], cannot by itself establish value because it concerns a different stage of care from the confirmatory tests it triggers. International economic evaluations are mixed: AI-supported mammography was cost-saving in Singapore [37] and cost-effective in Iran and Korea [38,39], but in the United Kingdom and Norway it was cost-effective only at low per-examination prices [40,41], and the international readiness review for the Philippines cautioned that fragmented information systems and unvalidated models may limit transferability [42].

Two distinct propositions are often conflated under the label “AI screening”: (i) applying an algorithm to examinations that the health system already acquires for other reasons, where the only new costs are the software and the work-up of additional positives, and (ii) building new screening programmes with AI-supported reading, where acquisition dominates the budget. This study separates them. Its objectives were to estimate, for breast, lung, colorectal and cervical cancer and from the perspective of the Philippine public payer: (1) the annual health gains, costs and incremental cost-effectiveness ratios (ICERs) of opportunistic AI reading and of an AI-supported screening programme; (2) the AI prices, specificities and acquisition costs at which each would meet Philippine cost-effectiveness benchmarks; (3) the five-year budget impact; (4) the expected value of further research, by parameter; and (5) the radiotherapy courses that auto-contouring could add within present staffing. All inputs are traceable to published sources or explicitly labelled as assumptions, and the code and data accompany the article so that the analysis can be updated as Philippine data accrue.

2. Materials and Methods

2.1. Study Design and Reporting

A deterministic and probabilistic decision-analytic model was built in Python 3.12 (NumPy 2.5, pandas 3.0, Matplotlib 3.11). Reporting follows the Consolidated Health Economic Evaluation Reporting Standards 2022 [43]; the completed checklist is provided as Supplementary Article S1. The reference case follows the Philippine HTA Methods Guide: a publicly funded health-care payer perspective (PhilHealth and the Department of Health), costs and outcomes discounted at 7% per year, and cost–utility analysis as the primary form of evaluation [44]. Because the Philippines has no explicit cost-effectiveness threshold [45], results are compared with 0.5× and 1× gross domestic product (GDP) per capita, the benchmarks used in recent HTAC assessments [18], and with the supply-side estimates of PHP 84,317 (mean) and PHP 209,411 (median) per QALY recently derived for the Philippines [46]. GDP per capita in 2024 was PHP 228,320 [47], giving thresholds of PHP 114,160 and PHP 228,320 per QALY. All costs are in 2024 Philippine pesos (PHP 57.29 = USD 1; [48]). The model takes an annual steady-state view: it estimates the costs and health gains attributable to one year of operation of each strategy, with life-years accruing over the remaining life expectancy of people whose deaths are averted. The study used only published aggregate data and required no ethical approval.

2.2. Model Structure

Figure 1 shows the model. For each cancer, incident cases are distributed across three summary stages, localized, regional and distant, the classification in which Philippine registry stage and survival data are reported. Two deployment routes feed a common stage-shift and prevention module and a budget and outcome module.

Route A, opportunistic AI reading. AI is applied to examinations already acquired for other reasons: chest radiographs taken in tuberculosis case-finding and for employment or clinical indications, diagnostic and opportunistic mammograms, cervical cytology, and colonoscopies. Each examination is already read by a human, so the AI pathway is credited only with the increment in sensitivity over human reading, ΔSn, and charged only with the increment in the false-positive rate, ΔFPR. The number of additional cancers detected in the year is

$$D_A=E_A\hspace{0.17em}{q}_A\hspace{0.17em}\Delta \mathit{Sn},$$

where E_A is the number of examinations routed through AI (annual volume × 80% coverage) and q_A the prevalence of detectable cancer per examination. Additional false positives are E_A(1 − q_A) ΔFPR. The only new cost per examination is the AI read.

Route B, AI-supported screening programme. Participation among the eligible population rises from the current 1% to 20%, and the newly screened are examined at the interval set by PhilHealth and HTAC rules (mammography every two years, low-dose or radiographic lung screening annually, colonoscopy every ten years, cervical cytology every three years). Annual examinations are E_B = N_elig(0.20 − 0.01)/I, where I is the screening interval in years. Because these people are otherwise unscreened, the full AI-supported sensitivity Sn applies, D_B = E_B q_B Sn, false positives are E_B(1 − q_B)(1 − Sp), and the payer bears the PhilHealth benefit rate for acquisition and reading plus the AI read.

Stage shift. Cancers detected through the AI pathway, D, capped at annual incidence N, follow a screen-detected stage distribution d concentrated in localized disease, while the remainder follow the status-quo distribution p: p' = (1 − s)p + sd with s = D/N. Expected treatment cost per case is p ⋅ c, where c is the stage-specific cost vector; deaths within five years are N p ⋅ (1 − S), where S is stage-specific five-year survival. Treatment cost avoided and deaths averted are the differences between the status quo and the shifted distribution. This interception logic is the one used in modelled evaluations of early-detection tests [49].

Prevention pathway. For colorectal and cervical cancer, most of the benefit of better reading arises from detecting precursor lesions, advanced adenomas and CIN2+, rather than invasive cancer. Additional precursors detected, P = E q_pre Sn_eff, are treated at a unit cost, and a proportion π would otherwise have progressed to invasive cancer after a mean lag L. Each prevented cancer saves the status-quo expected treatment cost and averts p ⋅ (1 − S) deaths, discounted by (1 + r)^−L.

Outcomes. Each averted death is credited with the discounted life-years from the mean time of death within the five-year window (2.5 years) to the remaining life expectancy at the mean Philippine age at diagnosis, taken from WHO life tables [50]. Life-years are weighted by a health-state utility of 0.80 to give QALYs. The ICER is incremental cost divided by incremental QALYs.

Radiotherapy module. Deliverable radiotherapy courses per year are the lesser of physician planning capacity (radiation oncologists × planning hours per year ÷ hours per course) and machine capacity (accelerators × courses per accelerator-year). Auto-contouring reduces the contouring share of planning time by 30% (conservative) or 76% (upper bound). Need is 53% of incident cancers, the optimal utilisation benchmark cited for the Philippines [11].

2.3. Parameters and Sources

Table 1 lists site-specific inputs and Table 2 global inputs; the full parameter file with provenance is provided as Supplementary Data S1 and the raw extracted inputs as Supplementary Data S2.

Incidence and mortality are GLOBOCAN 2024 estimates [1].

Stage at diagnosis. For breast cancer, the extent-of-disease distributions of the Metro Manila (1994–1995, n = 415) and Rizal (1996–1997, n = 1,299) registries were pooled after proportional redistribution of cases of unknown extent, giving 28% localized, 58% regional and 14% distant [4,51]; this is consistent with the 53% late-stage presentation reported for Philippine women in the ACTION cohort [19] and with the 5% stage I share in Metro Manila incident cases of 1993–2002 [76]. Colorectal cancer used the Manila and Rizal distribution of 1993–2002 (30% localized, 46% regional, 24% distant; [5]), and cervical cancer the Manila 1994–1995 distribution after redistribution of unknowns (34%, 49%, 17%; [4]). No population-based stage distribution exists for lung cancer; the Lung Center of the Philippines registry reported 66% stage IV among 3,969 patients in 2009–2017 [77] and 76% stage IIIB or higher in earlier admissions [78], and the residual was split between localized (12%) and regional (22%) disease in the proportions observed among Filipino-Americans in SEER [79]; this split is an assumption.

Screen-detected stage distribution. For breast cancer, 78% of cancers in the MASAI AI arm were stage 0–I and 22% stage II or higher [21], and in a Taiwanese series 67% of screen-detected cancers were stage 0–I and 8% stage III–IV, against 47% and 19% of non-screen-detected cancers [52]; a localized/regional/distant mix of 0.75/0.23/0.02 was used. Distributions for the other sites are assumptions.

Stage-specific survival. The reference case uses Philippine five-year survival by extent of disease from the Manila and Rizal registries: breast 68%/37%/10% (pooled), colorectal 64%/31% for localized and regional disease (colon and rectum pooled), and cervix 63%/30%/7% [4,51]. These are absolute survival proportions from the 1990s and therefore conservative. Distant colorectal survival, printed as 0% for colon and missing for rectum, was set to 11% by scaling SEER distant survival to the Philippine/US ratio of overall survival [5]. Lung cancer, not covered by the registry chapters, used SEER 2016–2022 stage-specific relative survival (65.5%/38.2%/10.5%) scaled to the Philippine/US ratio of overall relative survival, 12.0/17.3 [6,53], giving 45%/26%/7%. A scenario replaced all sites with contemporary SEER stage-specific relative survival.

AI accuracy. Breast: sensitivity 80.5% (95% CI 76.4–84.2) and specificity 98.5% for AI-supported reading, with an increment of 6.7 percentage points over double reading and 0.1 percentage point in the false-positive rate [21,22]. Lung: stand-alone sensitivity of 72% for lung cancer on radiographs [59,60,65]; programme-level specificity of 98%, consistent with a 1.3% flag rate in a 135,998-radiograph tuberculosis-programme pilot [64] and a 2.3% positive-report rate in the Korean trial [23]; an incremental detection of 30 percentage points over human reading of radiographs not read for nodules, from the doubling of actionable-nodule detection in that trial and reader-study gains of 11 points [23,65,66]; and an incremental positive rate of 0.4 points [23]. Colorectal: computer-aided detection raises advanced colorectal neoplasia detection by a relative 16% [25], applied to a baseline colonoscopy sensitivity of 0.90 (incremental 14 points), with 0.18 additional non-neoplastic polypectomies per colonoscopy [24] and per-polyp sensitivity 95% and specificity 88% [61]. Cervical: AI cytology sensitivity 90% and specificity 85% for primary screening [62,63], an incremental sensitivity of 5.8 points (95% CI 3.0–8.6) over manual reading [26], and an assumed 1 point increase in the false-positive rate.

Detectable prevalence. Lung cancer per opportunistic radiograph was 1.0 per 1,000, the yield in the Vietnamese tuberculosis-programme pilot [64], and 4 per 1,000 in a high-risk programme population (assumption). Breast cancer per examination was 6 per 1,000 (opportunistic) and 5 per 1,000 (programme), within the range of Asian programme detection rates of 2.7–6.4 per 1,000 [21,67,68,80]. Colorectal cancer per colonoscopy was 5 per 1,000 (opportunistic) and 8 per 1,000 (programme, reflecting the ten-year interval), with advanced neoplasia at 3.0% [69]. Invasive cervical cancer per cytology was 2 per 1,000 (opportunistic) and 1.5 per 1,000 (programme), with CIN2+ at 6 and 5 per 1,000, bracketed by 1.9 per 1,000 in a 1.47-million-woman Chinese programme, 4 per 1,000 in Thai primary HPV screening and 2.9% in African clinic populations [27,70,71]. Progression to invasive cancer was 30% for advanced adenomas over a five-year lag, between the 25% and 43% ten-year cumulative risks estimated from 840,149 screening colonoscopies [72], and 20% for CIN2+ over a ten-year lag, below the 31% thirty-year risk observed for untreated CIN3 because CIN2 frequently regresses [73].

Volumes and populations. No national counts of radiographs, mammograms, cytology tests or colonoscopies are published. Opportunistic volumes were therefore assumptions, bracketed by programme information: 3 million chest radiographs per year (Department of Health tuberculosis case-finding targets of one to twelve million people screened; Philippine Cancer Society programme of 25,000 [36]), 400,000 mammograms, 400,000 cytology tests and 150,000 colonoscopies, each varied from 50% to 150% in probabilistic analysis. Eligible populations were taken from the Philippine Statistics Authority 2024 projections: 14.79 million women aged 40–69 (mammography), 21.44 million women aged 30–64 (cervical), 18.35 million adults aged 50–74 (colorectal) and 25% of 19.48 million adults aged 50–79 assumed to meet smoking-history criteria (lung) [81].

Costs. Acquisition costs in the programme route are PhilHealth benefit rates: mammogram PHP 2,610 and screening colonoscopy PHP 23,640 [15]; chest radiograph PHP 370 from a bottom-up hospital costing [54]; and cytology PHP 800 (assumption). The AI read was PHP 250 per examination in the base case, between the PHP 50 quoted in the Philippines [36] and the PHP 270–610 implied by UK and Norwegian prices of GBP 4.72–10.62 and EUR 3 per examination [40,41]; it was varied between PHP 50 and 750 in sensitivity analyses. False-positive work-up used PhilHealth rates: core biopsy (PHP 3,640 adjusted by the 30% case-rate increase of 2024) plus breast ultrasound for breast (PHP 6,100), low-dose CT for lung (PHP 7,220), and assumptions of PHP 5,000 for polypectomy and histology and PHP 4,000 for colposcopy and biopsy [15,75,82]. Treatment cost of localized disease was anchored to published Philippine figures: PHP 350,000 for early breast cancer [54], the Z Benefit rates of PHP 150,000 for stage I–II colon cancer and PHP 175,000 for stage IA1–IIA1 cervical cancer [83,84], and an assumed PHP 400,000 for lung cancer. Stage-cost gradients relative to localized disease were 1.68 and 1.80 for breast, from the systematic review of Sun et al. [55] collapsed to summary stage; 2.00 and 2.05 for colorectal, from the Z Benefit tier for high-risk stage II–III disease (PHP 300,000) and a Malaysian stage IV:I ratio of 2.05 [56]; 2.20 and 2.80 for cervical cancer, from Chinese stage-specific costs [57]; and 1.30 and 1.10 for lung cancer, reflecting the near-flat gradient observed in China [58] and the Philippine pattern in which 57% of lung-cancer patients receive supportive care only [77]. A steep gradient of 1 : 2.57 : 4.00 (from stage I–IV ratios of 1 : 1.8 : 3.6 : 5.6) was examined as a scenario representing fuller treatment of advanced disease under the expanded breast-cancer Z Benefit package [85]. Radiotherapy course cost was 25 sessions at the adjusted case rate of PHP 3,900 plus PHP 18,000 planning (PHP 115,500) [74,75].

Life expectancy and utility. Remaining life expectancy at the mean Philippine age at diagnosis (51.4 years for breast, 59.9 lung, 58.8 colorectal, 49.8 cervix; [6]) was taken from WHO life tables for the Philippines, 2021 (24.1 years for women aged 51, 14.9 at 60 for both sexes, 15.6 at 59, 24.9 for women aged 50) [50]. No Philippine utility weights for cancer survivors were identified; the utility of 0.80 is an assumption varied between 0.70 and 0.90, and life-years are reported alongside QALYs.

Radiotherapy. Inputs were 125 radiation oncologists and 53 accelerators [10], 450 courses per accelerator-year (mid-point of 200–900 patients per centre; [11]), 500 planning hours per oncologist-year, four hours of physician time per course of which 60% is contouring (assumptions), and auto-contouring time savings of 30% [30] and 76% [28], with intermediate values reported by Radici et al. [29], Doolan et al. [86] and Moran, Poole & Barrett [87].

2.4. Analyses

Deterministic results were computed for both routes and for scenarios with a 3% discount rate, the steep cost gradient, SEER stage-specific survival and an AI price of PHP 50. Probabilistic sensitivity analysis drew 10,000 parameter sets per cancer and route with distributions chosen according to ISPOR-SMDM guidance [88]: beta for sensitivities, specificities, stage-specific survival (standard deviations 0.05, 0.05 and 0.03 for localized, regional and distant disease) and progression probabilities (standard deviation 0.06); normal distributions truncated at zero for the incremental sensitivity and incremental false-positive rate of the opportunistic route; Dirichlet for stage vectors (concentration equal to 60 times the point estimates); log-normal for costs (coefficients of variation 10–25%) and for cancer and precursor prevalence (35%); uniform for examination volumes (±50% opportunistic, ±20% programme) and utility (0.70–0.90); and, for the AI price, uniform between PHP 50 and 750 in the opportunistic route and log-normal around PHP 250 (standard deviation 0.4 on the log scale) in the programme route. Cost-effectiveness planes, acceptability curves [89] and probabilities of cost-effectiveness at both benchmarks were computed from net monetary benefit. One-way sensitivity analyses varied each parameter across its plausible range and ranked them by swing in the ICER. Threshold analyses identified the AI price and incremental false-positive rate at which opportunistic reading meets each benchmark and the maximum all-in cost per screening examination at which the programme meets them. Five-year budget impact was computed undiscounted with linear ramp-up to full coverage in year three [90]. The expected value of perfect information (EVPI) was calculated from the probabilistic sample for each site and route, and the expected value of partial perfect information (EVPPI) for individual parameters by the nonparametric regression method [91,92], expressed per year of national decision-making. The radiotherapy module was evaluated deterministically.

3. Results

3.1. Opportunistic AI Reading

At the base-case AI price of PHP 250, routing about 3.16 million already-acquired examinations through AI reading would detect about 970 additional cancers at an earlier stage, detect about 615 precursor lesions and thereby prevent about 173 cancers, and avert about 272 deaths within five years (161 through stage shift and 111 through prevention), gaining 1,845 discounted life-years and 1,476 QALYs per year (Table 3). The stage distributions move only slightly because the incremental detection reaches 0.5–5% of incident cases (Figure 2A). Net annual cost to the payer would be PHP 0.93 billion: PHP 0.79 billion for AI reads, PHP 0.19 billion for work-up of about 34,600 additional positives, offset by PHP 0.06 billion of treatment cost avoided (Figure 3A). The ICER for the four sites combined was PHP 628,000 per QALY, 2.8 times GDP per capita.

Results differed sharply by site. Computer-aided colonoscopy had the most favourable ratio, PHP 220,000 per QALY, because the prevention pathway contributes 97 of its 108 averted deaths and because the examinations already carry most of the cost; its probability of cost-effectiveness was 0.42 at 1× GDP and 0.17 at 0.5× GDP per capita (Table 4, Figure 4 and Figure 5A). Opportunistic AI mammography cost PHP 378,000 per QALY (probability 0.19 at 1× GDP): with incremental sensitivity of 6.7 points and a detectable prevalence of 6 per 1,000, 320,000 reads yield about 129 earlier diagnoses and 23 averted deaths. Chest-radiograph reading, despite the largest absolute gains (720 earlier diagnoses, 122 deaths averted), cost PHP 936,000 per QALY, because 2.4 million reads at PHP 250 cost PHP 0.60 billion and a near-flat stage-cost gradient leaves little treatment cost to recoup. AI cytology cost PHP 873,000 per QALY because invasive cancer and CIN2+ prevalence per cytology test are low. Across the probabilistic sample the four-site net cost ranged from PHP 0.50 to 2.74 billion (95% interval), and the probability that all four together met the 1× GDP benchmark was 0.01.

3.2. AI-Supported Screening Programme

Raising participation from 1% to 20% would generate about 4.04 million screening examinations per year once intervals are respected (1.41 million mammograms, 0.93 million lung examinations, 0.35 million colonoscopies and 1.36 million cytology tests), detect about 12,800 cancers (reaching 15–23% of incident cases, Figure 2B), detect about 16,000 precursor lesions and prevent about 4,200 cancers, and avert about 4,750 deaths per year, about 17 times the gain of opportunistic reading (Table 3). Net annual cost would be PHP 14.0 billion (95% interval 11.1–17.3 billion): PHP 14.5 billion of acquisition, reading and precursor treatment, PHP 1.3 billion of work-up for 284,000 false positives, and PHP 1.7 billion of treatment cost avoided (Figure 3B). The combined ICER was PHP 538,000 per QALY and the five-year budget impact PHP 56 billion (Figure S1); because the model is linear in uptake, deaths averted and net cost scale proportionally (1,000 deaths and PHP 3.0 billion at 5% uptake; 12,260 deaths and PHP 36.2 billion at 50%) while the ICER is unchanged.

Lung screening of the high-risk population by AI-read radiography was closest to the benchmark (PHP 264,000 per QALY; probability 0.36 at 1× GDP), owing to low acquisition cost (PHP 370) and high distant-stage mortality in the status quo. Cervical cytology with AI reading cost PHP 373,000 per QALY (probability 0.14), AI-supported mammography PHP 448,000 (probability 0.10) and screening colonoscopy PHP 758,000 (probability 0.004), the last because the PHP 23,640 benefit rate dominates (Table 4; Figure 5B and Figure S2). No site had a material probability of being cost-saving in either route.

3.3. Scenario, One-Way and Threshold Analyses

Table 5 summarises the scenarios. Discounting at 3% instead of 7% reduced the four-site ICERs to PHP 408,000 (opportunistic) and PHP 320,000 (programme) per QALY. The steep stage-cost gradient reduced them to PHP 341,000 and PHP 421,000, and the opportunistic net cost from PHP 0.93 to 0.50 billion, because down-staging then recoups more treatment expenditure. Contemporary SEER stage-specific survival changed the opportunistic ICER little (PHP 608,000) but worsened the programme ICER (PHP 706,000), because the survival advantage of localized over regional disease is narrower in modern series than in the Philippine registries of the 1990s. An AI price of PHP 50 reduced the opportunistic ICER to PHP 200,000 per QALY, below the 1× GDP benchmark, while leaving the programme ICER almost unchanged (PHP 507,000).

One-way analyses (Figure 6; Figure S3) rank AI price first for opportunistic chest-radiograph reading, with the ICER spanning PHP 261,000 to 2.62 million per QALY between PHP 50 and 750 per read, followed by incremental sensitivity (PHP 609,000 to 2.01 million), detectable prevalence, the stage-cost gradient and stage-specific survival; examination volume does not change the ratio because costs and benefits scale together. For the mammography programme, detectable prevalence (PHP 270,000 to 980,000 per QALY), acquisition cost (PHP 306,000 to 589,000), survival, the discount rate and the cost gradient dominate, while AI sensitivity and specificity matter little once acquisition is paid for.

Threshold analyses (Figure 7) locate the conditions for cost-effectiveness. Opportunistic reading of all four examination types together meets 1× GDP per capita only at an AI price of PHP 50 or less per read; AI mammography meets it at PHP 150 or less (PHP 100 at 0.5× GDP), computer-aided colonoscopy at PHP 275 or less, and chest-radiograph and cytology reading only at PHP 25 or less. For chest radiographs the incremental false-positive rate is a secondary lever: every additional percentage point of positives adds about PHP 240,000 per QALY to the ICER at current CT costs. For the programme, the maximum all-in cost per examination (acquisition plus AI read) consistent with 1× GDP per capita is PHP 1,645 for mammography against a current PHP 2,860, PHP 519 for a lung radiograph against PHP 620, PHP 8,060 for colonoscopy against PHP 23,890, and PHP 489 for cytology against PHP 1,050; at 0.5× GDP the limits fall to PHP 1,013, 195, 4,648 and 48.

3.4. Value of Information

At 1× GDP per capita the EVPI per year of decision-making was PHP 37.7 million for computer-aided colonoscopy, PHP 7.1 million for opportunistic AI mammography and under PHP 2.1 million for radiograph and cytology reading, where the decision is not close (Figure 8A). For the programme (not shown in Figure 8), EVPI was PHP 110 million for lung screening, PHP 96 million for mammography, PHP 69 million for cervical screening and PHP 4.6 million for colonoscopy. Among opportunistic uses, the parameters of greatest partial value were incremental sensitivity over human reading (PHP 25.6 million per year, almost entirely for colonoscopy), precursor prevalence per examination (PHP 15.3 million), AI price (PHP 10.6 million), precursor progression probability, false-positive work-up cost and the incremental false-positive rate (PHP 6.3–7.4 million each) (Figure 8B). For the programme, detectable prevalence per screening examination dominated (PHP 115 million per year), followed by the distant-stage cost gradient (PHP 29 million) and AI price (PHP 10 million). Examination volume, treatment cost of localized disease and stage-specific survival carried little value, because they shift costs and benefits together or because the decision does not turn on them.

3.5. Radiotherapy Planning Capacity

With 125 radiation oncologists each devoting 500 hours per year to planning at four hours per course, physician time limits delivery to about 15,600 courses per year, below the machine ceiling of about 23,900 courses and far below the roughly 79,400 courses implied by 53% utilisation of 149,852 incident cancers (Table 6; Figure 9). Auto-contouring that removes 30% of delineation time raises the planning ceiling to about 19,100 courses (+3,430 per year); at the 76% upper bound the planning ceiling (28,700) exceeds machine capacity, which then binds at 23,900 courses (+8,225 per year). The corresponding additional treatment expenditure at PhilHealth rates is PHP 0.40–0.95 billion per year. Beyond that point, further gains require accelerators and staff rather than software.

4. Discussion

4.1. Principal Findings

Under the Philippine HTA reference case, neither route to AI-supported cancer detection is self-financing, and cost-effectiveness depends on three conditions that the analysis makes explicit. First, the price of the software: opportunistic AI reading of examinations already acquired meets the 1× GDP benchmark only at or below PHP 50 per read for the four sites together, which is the price reported for the Philippine Cancer Society’s chest-radiograph programme [36] but one-fifth to one-tenth of prices observed in European screening services [40,41]. Second, the mechanism of benefit: computer-aided colonoscopy is the only opportunistic use that approaches the benchmark at the base-case price, because it prevents cancers by finding advanced adenomas rather than merely shifting the stage of cancers that would be found anyway. Third, the cost of acquisition: once a new screening examination must be purchased, AI accuracy becomes a minor driver, and the programme is cost-effective only if the all-in cost per examination falls well below current PhilHealth benefit rates, by about 40% for mammography and 50% for cytology, and by two-thirds for colonoscopy.

These conclusions differ from an earlier modelling exercise by the author, which credited the AI pathway with its full sensitivity and so found opportunistic triage cost-effective at a much lower ICER. The present model is more conservative in three respects that matter: it credits AI only with its incremental yield over the human read that already takes place, it applies the 7% discount rate of the Philippine HTA Methods Guide, and it uses Philippine registry survival. The results are, however, concordant with the international literature once price is taken into account. In the United Kingdom, AI as an additional reader was cost-effective only at a price of about GBP 5.50 per scan and not at the GBP 10.62 cost of AI reading in the national programme [40]; in Norway the break-even running cost was about EUR 1 per examination [41]; and the favourable Singaporean, Iranian and Korean analyses either modelled AI replacing a human reader, so that reading cost was saved, or assumed low software costs within lifetime Markov models [37,38,39]. A Vietnamese budget-impact analysis of incidental nodule detection on tuberculosis-programme radiographs reported cost neutrality only by year five [93]. The Philippine payer currently reimburses reading within the examination fee, so no reading cost is saved by AI unless the benefit rate is renegotiated; the programme-route results therefore show what AI can and cannot do when it is purely additive.

4.2. Implications for Policy

The findings suggest a sequence rather than a verdict. For PhilHealth, the immediate implication is that software used within reimbursed screening pathways should be financed under evidence-development conditions, as proposed in the companion review [31]: coverage conditional on reporting of flag rates, confirmatory-test completion, cancers detected per confirmatory test and stage, so that the parameters of highest value of information, incremental yield over the human read, detectable prevalence per examination and false-positive rates, are measured in Philippine practice. The value-of-information analysis gives these data a price: about PHP 38 million per year for computer-aided colonoscopy and about PHP 100 million per year each for the lung and breast programmes, which would justify prospective evaluations of modest size, for example within the Philippine Cancer Society cohort or in facilities accredited under Circular 2025-0027.

For procurement, the threshold analysis converts cost-effectiveness into a maximum reimbursable price per read, as the UK analysis did [40]: about PHP 150 for mammography, PHP 275 for colonoscopy software, and PHP 25–50 for chest radiographs and cytology at 1× GDP per capita, lower at stricter thresholds. Prices in this range are plausible at national volume (the Philippine chest-radiograph quotation is PHP 50, and programme costs of USD 6.31 per woman have been reported for AI-assisted cytology in China; [70]), but they should be negotiated with these figures in hand. For chest radiography, the analysis also shows that deployment should be tuned to a high-specificity operating point and paired with a funded CT pathway, because each additional percentage point of positives adds about PHP 240,000 per QALY at the current PHP 7,220 CT benefit; the Philippine tuberculosis programme’s experience of threshold calibration is directly relevant [35].

For screening policy, the programme results confirm that screening buys health rather than money [19]: a 20% programme would avert about 4,750 deaths per year, seventeen times more than opportunistic reading, but at PHP 14 billion per year. Within it, radiograph-based lung screening of high-risk adults is closest to cost-effectiveness, and its ICER would fall further if low-dose CT were reserved for AI-flagged radiographs rather than used as the first-line test; this two-step design, consistent with Circular 2025-0027 and with the proposal of the Philippine Cancer Society, deserves a dedicated analysis. For colorectal cancer, screening colonoscopy at the current benefit rate cannot be cost-effective as a primary test, which supports the HTAC’s ongoing assessment of faecal immunochemical testing as the entry point, with computer-aided detection reserved for the colonoscopies that follow. For cervical cancer, the HTAC’s recommendation of HPV DNA testing [18] means that AI cytology or automated visual evaluation would serve as triage rather than primary screening, a role in which specificity matters more than the sensitivity gain modelled here [27]. The scenario analyses also show that fuller treatment of advanced disease, as the expanded breast-cancer Z Benefit package intends [85], strengthens the economic case for early detection by steepening the stage-cost gradient.

The radiotherapy module points to a use of AI that relieves the scarcer input directly. Physician planning time, not machine time, appears to cap delivery at about 15,600 courses per year, and auto-contouring could raise this by roughly a fifth to a half within present staffing before accelerators bind; the additional PHP 0.4–0.95 billion per year is treatment expenditure that PhilHealth would already owe under its case rates. The software itself is cheap relative to this expenditure, but the module does not value it formally and its inputs on planning hours are assumptions that the centres operating the 53 accelerators could readily replace with logged data.

To the best of our knowledge, this is the first Philippine economic evaluation to jointly assess the cost-effectiveness, budget impact and value of information of AI-supported detection across breast, lung, colorectal and cervical cancer. Its principal contribution is the explicit separation of opportunistic AI reading from new screening programmes, distinguishing the incremental value of interpretive software from the costs of expanding screening. The analysis integrates Philippine population-based stage and survival data, PhilHealth reimbursement rates, and trial and meta-analytic evidence on AI performance within the Philippine HTA reference case. Crediting opportunistic AI only with its incremental yield over existing human reading strengthens the relevance of the comparisons. Pricing and acquisition-cost thresholds translate the findings into actionable financing conditions, while value-of-information analyses identify priorities for local evidence generation. Publicly provided code and inputs support transparency, reproducibility and future updating.

These findings should be interpreted within the model’s scope. The static annual structure does not explicitly capture lead time, overdiagnosis or repeated screening, and mortality outcomes are limited to five years after diagnosis. Philippine stage and survival inputs derive from historical Metro Manila and Rizal registries, with lung-stage data from a single hospital. Several inputs including examination volumes, prevalence, selected stage distributions, utility and unit costs are assumptions examined in sensitivity analyses. AI performance is drawn largely from international studies; Philippine validation and measurement of incremental yield remain priorities, particularly for chest radiography [31]. Implementation, indirect and household costs are excluded, and programme estimates assume access to screening at modelled rates and adequate confirmatory capacity. Cost-effectiveness also depends on the benchmark adopted, given the absence of an explicit Philippine threshold [46]. These qualifications define the conditions for applying the results and reinforce the value of the study’s threshold and research-priority analyses.

4.3. Future Research

The analysis identifies what to measure. Prospective evaluations embedded in financed pathways should record, for each examination type, the flag rate, confirmatory-test completion, cancers and precursors detected per confirmatory test, and stage, so that detectable prevalence, incremental yield and false-positive rates can be estimated for Filipino populations; these are the parameters with the highest value of information. Logged planning times in radiotherapy centres would replace the capacity module’s assumptions. A lifetime microsimulation calibrated to the registry, once the national registry reports, would supersede the present static structure. Finally, the price elasticity of software at national volume should be established through procurement dialogue, because the price per read is the single parameter that determines whether opportunistic AI reading is worth public money in the Philippines.

5. Conclusions

For the four cancers that account for about half of the Philippine cancer burden, AI-supported detection would avert deaths but would not pay for itself under current treatment patterns and PhilHealth rates. Opportunistic AI reading of examinations already acquired is cost-effective at the 1× GDP benchmark only when software is priced at or below about PHP 50 per read overall, up to PHP 150 for mammography and PHP 275 for computer-aided colonoscopy, whose prevention of cancers through adenoma detection gives it the most favourable ratio. An AI-supported screening programme would avert about 4,750 deaths per year at PHP 14 billion per year and would require all-in costs per examination well below current benefit rates; AI-read radiography of high-risk adults is the component closest to cost-effectiveness. Auto-contouring could add 3,400–8,200 radiotherapy courses per year within present staffing. The parameters that most warrant Philippine data are incremental yield over human reading, detectable prevalence per examination and the price of the software; collecting them through evidence-development conditions on reimbursed software would allow the Department of Health and PhilHealth to decide on firmer ground.