Abstract
Background. High-dose vitamin C has been promoted as a cancer treatment since the 1970s. Oral and intravenous (IV) dosing produce very different blood levels, and the two routes have often been discussed as if they were one intervention. A 2025 meta-analysis pooled randomised trials with cohort studies, used a median survival ratio, and concluded that IV vitamin C had moderate-certainty evidence and should be considered as standard of care. Two randomised trials published around that analysis point in opposite directions, and none of its included cancer types had more than one randomised trial.
Objectives. To assess the effects of high-dose vitamin C, given orally or intravenously, on overall survival, progression-free survival, response, quality of life and harms in adults with cancer, with randomised evidence kept separate from non-randomised evidence and each cancer type judged on its own data.
Search methods. PubMed/MEDLINE via E-utilities (620 records, all retrieved), Europe PMC with full cursor pagination (8,521 records, all retrieved) and ClinicalTrials.gov (56 records) were searched on 13 November 2026, with no date or language limits in the strings as run. Embase, Scopus, Web of Science, CENTRAL and CINAHL were not run (no access) and are listed with ready strings in the search log. The WHO ICTRP portal could not be searched through an API and is listed as not run. OpenAlex was rate-limited (HTTP 429) on two attempts and is logged as blocked. Citation chasing was carried out on the 2025 competing meta-analysis, the 2026 J Med Life review, the NCI PDQ summary and the included studies.
Selection criteria. Comparative studies in adults with confirmed cancer: randomised trials for the primary pool, non-randomised comparative studies analysed separately. IV stratum: at least 3 g/day or at least 0.5 g/kg per infusion. Oral stratum: at least 3 g/day. The strata were never pooled. Single-arm studies, case series and prevention studies in people without cancer were excluded from the pools.
Data collection and analysis. Searching, screening, extraction and risk of bias assessment were carried out by AI agents under the author's direction. Two independent screening passes (agreement 94.7%, Cohen kappa 0.61, moderate agreement on the three-level decision and 0.70 on the binary decision) with adjudication by reading of all 800 records either pass included. Two fully independent extractions (critical-field discrepancy rate 6.4%), with every critical discrepancy settled by re-reading the source. RoB 2 and ROBINS-I in duplicate (2 of 18 overall judgments disagreed, both settled by an adjudicator reading the papers). Random-effects meta-analysis (REML) with Hartung-Knapp intervals, tau-squared, I-squared and prediction intervals. GRADE per outcome and per cancer type.
Main results. 18 studies met the inclusion criteria: 10 randomised trials, 1 comparative study whose randomisation is not described in its report (Zhao 2018), and 7 non-randomised comparative studies. Searching, screening and extraction were carried out by AI agents under the author's direction, with every stage done twice and disagreements settled by reading. Five databases were not searched and two records could not be classified; both limits are carried into the conclusions. Four IV trials reported an overall survival hazard ratio, one per cancer type. Colorectal cancer (Wang 2022, phase III, n=442): HR 1.04 (95% CI 0.81 to 1.33), moderate certainty. Pancreatic cancer (Bodeker 2024, phase II, n=34 treated, stopped early): HR 0.46 (90% CI 0.23 to 0.92; expressed at 95% that interval is about 0.20 to 1.05, crossing no effect), very low certainty. Non-small cell lung cancer (Ou 2020, n=97, vitamin C always given with electro-hyperthermia): HR 0.33 (as printed, 95% CI 0.16 to 0.41), very low certainty for vitamin C itself. Prostate cancer (Paller 2024, n=47 treated, stopped for futility): HR 1.98 (95% CI 0.85 to 4.58), low certainty. The all-cancer IV pool gave HR 0.73 with I-squared 87.5% and a 95% prediction interval of 0.02 to 25.07, so it does not describe a common effect. Excluding the hyperthermia trial moved the pool to HR 0.99 (Wald 95% CI 0.49 to 2.01; Hartung-Knapp 0.19 to 5.26). A sensitivity pool of the two trials rated low risk of bias, Wang 2022 and Paller 2024, gave HR 1.26 (Wald 95% CI 0.71 to 2.24; Hartung-Knapp 95% CI 0.03 to 52.8, I-squared 51.6%), too imprecise at two studies to carry a conclusion on its own; the trial-level results are the headline. Progression-free survival in three trials pooled to HR 0.86 (Hartung-Knapp 95% CI 0.49 to 1.53), low certainty. Grade 3 or worse adverse events showed no large increase in screened trial populations (Wang 2022 risk ratio 1.10, 95% CI 0.84 to 1.45), low certainty, and no trial reported the named specific harms (haemolysis, oxalate kidney injury, stones, glucose meter errors), which the trials could not detect well because they screened out the at-risk patients. The two non-randomised studies with adjusted hazard ratios were not pooled: one is a transplant repletion cohort (adjusted HR 0.50) and the other a breast cancer cohort in which vitamin C was given with mistletoe (adjusted HR 0.561). Each is very low certainty on its own. Oral vitamin C trials from 1979 and 1985 reported no survival advantage and report no hazard ratios. A descriptive replication of the median-ratio method on randomised trials only gave a geometric mean ratio of 1.15 (range 0.52 to 1.93), against the 1.83 reported in 2025 from a mixed pool.
Authors' conclusions. The randomised record does not support the 2025 claim of moderate-certainty benefit. The largest trial found no significant survival benefit; its interval runs from a 19% hazard reduction to a 33% increase, so it does not prove there is none. One small pancreatic trial found a large benefit and was stopped early, and early stopping tends to flatter a result. Five databases were not searched, so these conclusions are bounded by the searches actually run. High-dose IV vitamin C remains investigational. The US Food and Drug Administration has not approved it as a treatment for cancer.
Keywords: ascorbic acid; vitamin C; neoplasms; survival; systematic review; meta-analysis
Plain language summary
Vitamin C at normal doses is a nutrient. At very high doses given into a vein, blood levels rise far above anything tablets can reach, and laboratory work shows those levels can damage cancer cells. That laboratory finding has driven fifty years of hope and argument.
We gathered the comparative studies our search found (three databases in full, with five more not searched, a limit stated in the main text), kept tablet studies separate from vein studies, and kept randomised studies separate from weaker designs. Eighteen studies qualified. For survival, the randomised evidence is one trial per cancer type, and the trials disagree. A large bowel cancer trial with 442 people found no benefit. A small pancreatic cancer trial with 34 treated patients reported that people given vitamin C lived a median of 16 months against 8.3 months without it. That sounds large, the trial was stopped early because the result looked good, and early stopping tends to flatter a result. The same trial's own statistics, expressed the usual way, are also compatible with no benefit at all. It is a lead for a bigger trial, not an answer. A lung cancer trial combined vitamin C with heat treatment, so its result cannot be credited to vitamin C alone. A prostate cancer trial found no benefit and was stopped because it was not working.
The two trials with the soundest methods did not show that vitamin C helped people live longer. One of them, in bowel cancer, found the two groups lived about the same length of time. The other, in prostate cancer, found the group given vitamin C lived a shorter time on average, but the trial was small, so that difference could be chance. Vitamin C by tablet was tested in two careful blinded trials in 1979 and 1985 and did not help people live longer. In one breast cancer trial, people given vitamin C by vein for four weeks reported less nausea, better appetite, less pain, less fatigue and better sleep. An earlier review combined randomised and non-randomised studies and reported a much rosier survival figure than the randomised trials support. The difference comes from which studies are mixed together. In almost every trial in this review, vitamin C was added to standard cancer treatment. No trial compared vitamin C instead of standard treatment in a way that could support choosing it over proven therapy.
Vitamin C by vein did not appear to add much serious side effect burden in these trials. That statement comes with a condition worth holding on to: the trials screened out the people most at risk before the first infusion. People with G6PD deficiency, a red blood cell enzyme condition, can have their red blood cells break down after high-dose vitamin C. People with kidney disease can develop a form of kidney injury from it, or kidney stones. It can also make some finger-prick glucose readings read too high, which matters for people using insulin. If any of those apply to you, they are a reason to raise them with your cancer team before any infusion, inside or outside a trial. Nothing in this review tells any individual person whether vitamin C would help or harm them. The trials in this review gave it on top of standard cancer treatment, in hospitals, with safety screening first. People considering paying privately for infusions are buying an unproven treatment, and the known harms fall on exactly the people the trials screened out. High-dose vitamin C for cancer is investigational and has not been approved by the FDA.
Key findings
- 18 studies included: 10 randomised trials, 1 comparative study whose randomisation is not described in its report, and 7 non-randomised comparative studies.
- Four intravenous trials reported an overall survival hazard ratio, one per cancer type; no cancer type had more than one randomised trial.
- Colorectal cancer (Wang 2022, phase III, n=442): HR 1.04 (95% CI 0.81 to 1.33), moderate certainty. Prostate cancer (Paller 2024): HR 1.98 (0.85 to 4.58), low certainty.
- The all-cancer intravenous pool gave HR 0.73 with I-squared 87.5% and a 95% prediction interval of 0.02 to 25.07, so it does not describe a common effect.
- Progression-free survival in three trials pooled to HR 0.86 (Hartung-Knapp 95% CI 0.49 to 1.53), low certainty.
Summary of findings
Question: IV high-dose vitamin C plus standard therapy compared with standard therapy alone, in adults with cancer. Setting: trial settings, screened populations.
| Outcome | Studies (participants) | Effect | Certainty | Comment |
|---|---|---|---|---|
| Overall survival, colorectal | 1 RCT (442 randomised) | HR 1.04 (0.81 to 1.33) | Moderate | Phase III, low risk of bias. Any effect is likely small. |
| Overall survival, pancreatic | 1 RCT (34 treated) | HR 0.46 (90% CI 0.23 to 0.92) | Very low | Stopped early. At 95% the interval is about 0.20 to 1.05. |
| Overall survival, NSCLC | 1 RCT (97 randomised) | HR 0.33 (as printed, 0.16 to 0.41) | Very low | Vitamin C always combined with electro-hyperthermia. |
| Overall survival, prostate | 1 RCT (47 treated) | HR 1.98 (0.85 to 4.58) | Low | Stopped for futility. Point estimate favours control. |
| Overall survival, all IV RCTs pooled | 4 RCTs (625 randomised) | HR 0.73 (HK 0.21 to 2.60), I2 87.5%, PI 0.02 to 25.07 | Very low | Pool does not describe a common effect. |
| Overall survival, low-bias RCTs only (sensitivity) | 2 RCTs (492 randomised) | HR 1.26 (Wald 0.71 to 2.24; HK 0.03 to 52.8), I2 51.6% | Low | Sensitivity only. At two studies the HK interval is uninformative. |
| Progression-free survival, IV RCTs | 3 RCTs | HR 0.86 (HK 0.49 to 1.53) | Low | Ou 2020 excluded: printed CI invalid. |
| Overall survival, non-randomised | Simmons 2024 (110); Lee 2026 (748) | Adjusted HR 0.50 and 0.561, reported separately, not pooled | Very low each | Transplant repletion cohort; breast cohort with mistletoe co-therapy. Clinically unrelated, so pooling was withdrawn at peer review. |
| Grade 3+ adverse events | 3 RCTs | No large increase seen in screened populations | Low | Wang RR 1.10 (0.84 to 1.45) from counts 74/221 vs 67/221; a relative increase of up to about 45% is not excluded. |
| Specific harms (haemolysis, kidney injury, stones, glucose meter errors) | 0 RCT event reports | Not estimable | Very low | Trials screened out the at-risk groups. Absence of reports is a gap. |
| Quality of life / symptoms | 1 RCT (350), 1 cohort, 1 RCT no-detriment report | Short-term symptom benefit in one breast cancer trial | Low | Scales and constructs differ; not pooled. Downgrades: risk of bias (single-blind, self-reported outcomes) and single-trial imprecision per construct. |
| Overall survival, ovarian | 1 RCT (22 in survival set) | No HR reported; survival in graph form only | Very low | Ma 2014, not powered for efficacy. |
| Overall survival, oral stratum | 2 RCTs (250 randomised) | No HR reported; both trials report no advantage | Low | Creagan 1979 and Moertel 1985, double-blind, consistent null wording. |
1 Background
Cameron and Pauling reported in the 1970s that terminal cancer patients given supplemental ascorbate lived far longer than historical controls. Two double-blind Mayo Clinic trials then tested oral vitamin C at 10 g/day and found no benefit. Interest returned when pharmacology work showed why the route matters: the gut limits how much vitamin C blood can carry after a tablet, while an infusion bypasses that limit and reaches millimolar levels. At those levels ascorbate behaves less like a vitamin and more like a drug, generating hydrogen peroxide in tissue fluid.
Guideline and reference sources treat the question as open and investigational. The NCI PDQ summary on intravenous vitamin C (version dated 13 May 2025) records promising early data, general tolerability in trials, and limits in trial design, and states that the FDA has not approved high-dose vitamin C as a treatment for cancer. Memorial Sloan Kettering's monograph classifies high-dose IV ascorbate as an investigational drug for trial settings.
Three published syntheses frame the gap this review fills. Qu and colleagues (2025) pooled 8 studies (2,722 participants), mixing randomised trials with cohorts and pooling median survival, graded the body of evidence as moderate certainty, and concluded in their abstract that the evidence supports considering IV vitamin C as a standard of care in managing malignant neoplasms. That conclusion is the one the present review disputes, on the same data plus two trials. Shahzad and colleagues (2026) pooled single-arm and comparative reports by survival time with no comparator hazard ratio. Alangari and colleagues (2026) reviewed high-dose IV vitamin C across sepsis, oncology and symptom care, ran no meta-analysis, and called the oncology role supportive and exploratory. None of the three separated randomised from non-randomised evidence for a hazard-ratio analysis by cancer type, and the Qu search closed on 13 October 2024; neither the Bodeker pancreatic trial nor the Paller prostate trial was in its randomised set.
2 Objectives
To assess the effect of high-dose vitamin C (IV or oral) added to standard care, or given in place of it where no standard care was offered, on overall survival (primary), progression-free survival, objective response, quality of life and harms, in adults with cancer. Pre-specified questions, fixed in protocol v1: does the answer change when randomised and non-randomised evidence are kept apart; does it change by cancer type; and how much of the 2025 pooled median-ratio result survives a randomised-only, hazard-ratio analysis?
3 Methods
Methods follow the Cochrane Handbook (version 6.5) as applied in this series' Robust Standard v2. The full protocol, search log, screening records, both extraction passes, settlement record, risk of bias files, analysis code and datasets are held with this review and are listed in the data availability statement.
Registration and protocol. "Cochrane-level" in the title names this series' standard (protocol fixed before screening, dual independent screening and extraction, risk of bias in duplicate, GRADE per outcome); it does not claim Cochrane publication or endorsement, and this review's searches were not complete, as the Limitations state. This review was not registered on PROSPERO before analysis, and no registration number is claimed anywhere in this manuscript. The protocol was fixed on 13 November 2026 before screening; two dated amendments were logged before extraction (a competitor citation correction with a new competitor added, and citation-chasing additions with two reinstatements) and one at full-text deduplication. The amendments are in the review files and none changed the PICO, the outcomes or the subgroups.
Eligibility. Adults with histologically or clinically confirmed cancer in comparative studies. IV stratum: at least 3 g/day or at least 0.5 g/kg per infusion. Oral stratum: at least 3 g/day. Comparator: the same standard therapy with placebo, the same therapy alone, or best supportive care alone. Primary outcome: overall survival as a hazard ratio. Secondary: progression-free survival, objective response, quality of life on validated scales, grade 3 or worse adverse events, serious adverse events, and named specific harms (haemolysis, oxalate nephropathy or acute kidney injury, kidney stones, falsely raised point-of-care glucose readings, discontinuation for adverse events). Single-arm studies and case series were logged in a signals table and never pooled. Prevention studies in people without cancer were excluded.
Search. Reported in the Stage 1 search log with strings as run. PubMed/MEDLINE, Europe PMC and ClinicalTrials.gov were run in full on 13 November 2026 (9,197 records). Embase, Scopus, Web of Science, CENTRAL and CINAHL were not run; ready strings are in the log. WHO ICTRP has no usable search API from this environment (two routes tried) and is recorded as not run. OpenAlex returned HTTP 429 on two attempts and is recorded as blocked by rate limit; no further calls were made to that provider in this run. Citation chasing on Qu 2025, Shahzad 2026, Alangari 2026, the NCI PDQ summary and all included studies added two studies the database search had missed (Cameron and Campbell 1991; Zhao 2018) and surfaced two screening errors, corrected in a dated amendment before extraction (Ma 2014; Ou 2023).
Screening. Every deduplicated record (8,818) went through two screening passes built from two different written rule sets (design-first and outcome-first). Agreement was 94.7% and Cohen kappa was 0.61 on the three-level decision (0.70 on the binary decision). All 800 records that either pass included or marked maybe were read again by adjudicators, who recorded their own decision and reason for each record. That stage sent 45 records (35 studies) to full-text assessment and flagged 2 records for the author (R1550 and R2607: no abstract, and the title alone cannot settle eligibility; they remain unclassified, listed in the supplement). Full-text assessment excluded 30 of the 45 records (trial records with no results, duplicates of published reports, duplicate records, dose below threshold, single-arm reports, a superseded conference abstract, a letter with no data) and included 15 studies; chasing added two studies the search had missed, one screening error was corrected (Ma 2014), and one reinstated record (Ou 2023) turned out to be the definitive report of a study already in the set, which brought the final count to 18 studies.
Extraction and risk of bias. Two extractors worked apart, each extracting all 18 studies in full. A field-by-field merge compared 702 cells: the raw wording-inclusive discrepancy rate was 29.8%, and the discrepancy rate on the 360 critical analysis fields (effect estimates, intervals, denominators, route, dose, identifiers) was 6.4% (23 cells). Every critical discrepancy was settled by re-reading the source; in each of the six contested items the settlement is documented (settlement.md). Risk of bias used RoB 2 for randomised trials and ROBINS-I for the rest, in duplicate. Overall judgments agreed in 16 of 18 studies; an adjudicator settled the other two by reading the papers (Paller 2024: Low; Simmons 2024: Serious).
Analysis. Random-effects meta-analysis on log hazard ratios, REML estimator, Hartung-Knapp intervals, with tau-squared, I-squared and 95% prediction intervals. Prediction intervals are reported for the primary pool (0.02 to 25.07), the pool without the hyperthermia trial (0.0003 to 3591.6) and the progression-free survival pool (0.25 to 2.97). For a two-study pool a prediction interval is not defined, so none is reported for the low-bias sensitivity pool. Tau-squared is weakly identified at these study counts: a Q-profile 95% interval for the primary pool runs from 0.13 to 9.03 around the estimate of 0.519, and for the progression-free survival pool from 0 to 12.9 around a boundary estimate of 0. Estimator sensitivity was checked with DerSimonian-Laird (0.482 on the primary pool). A cancer-type stratum with fewer than two randomised trials reporting a hazard ratio was not pooled. Cohorts were pooled, if at all, apart from trials. Oral and IV strata were never pooled. Standard errors for Bodeker 2024 were taken from its reported 90% intervals. Ou 2020 was excluded from the progression-free survival pool because the confidence interval printed in that paper has its upper bound below its own point estimate, so no valid standard error exists. For overall survival, Ou's standard error was computed from the full width of its printed 95% interval on the log scale: (ln 0.4105 minus ln 0.1582) / 3.92 = 0.243. That interval is asymmetric around the estimate on the log scale, so bound-based alternatives differ (0.370 from the lower bound, 0.116 from the upper bound); the primary pool is therefore reported with and without Ou, and the without-Ou result does not depend on this choice. The statistician ran a fixed-effect weighting check on the primary pool's log inputs (weights Wang 69.1%, Ou 18.7%, Bodeker 6.2%, Paller 6.0%) and checked the crude participant totals of every pool against the extraction sheet; at peer review the participant total of the primary pool was corrected from 685 to 625, an error the first check had missed because it summed the same wrong figure. A descriptive replication of the median survival ratio method was run as a sensitivity analysis. GRADE was applied per outcome and per cancer type.
4 Results
Description of studies
18 studies: 10 randomised trials (Creagan 1979, Moertel 1985, Pathak 2005, Ma 2014, Ou 2020, Mansoor 2021, Wang 2022, van Gorkom 2022, Bodeker 2024, Paller 2024), 1 comparative study whose randomisation is not described in its report (Zhao 2018, counted as a randomised trial in the first draft and re-described at peer review), and 7 non-randomised comparative studies (Cameron and Pauling 1976, Cameron and Campbell 1991, Vollbracht 2011, Ou 2020 triple-negative breast, Ou 2023 NSCLC biomarker cohort, Simmons 2024, Lee 2026). Cancer types covered: colorectal, pancreatic, lung, prostate, ovarian, breast, acute myeloid leukaemia, mixed advanced cancers, and transplant populations with myeloma, lymphoma or other haematological disease. PRISMA counts: 9,197 records identified from databases, 8,818 after deduplication, 8,018 excluded by both screening passes, 800 adjudicated by reading (753 excluded, 2 flagged for the author, 45 sent on), 45 full-text records assessed (35 studies), 30 records excluded with reasons (trial records with no results 10, duplicates of published reports 9, duplicate records 6, dose below threshold 2, single-arm 1, superseded abstract 1, letter with no data 1), 15 studies included from that stage, plus 2 studies added by citation chasing (Cameron and Campbell 1991; Zhao 2018) and 1 study reinstated after a screening error was caught by chasing (Ma 2014), giving 18 studies. A characteristics table for all 18 studies is held in the supplement (extraction_final.csv) and summarised in the review files. The PRISMA diagram data are in the supplement.
Risk of bias
RoB 2 overall: Low 2 (Wang 2022, Paller 2024), Some concerns 5 (Creagan 1979, Moertel 1985, Pathak 2005, Mansoor 2021, van Gorkom 2022), High 4 (Ma 2014, Zhao 2018, Ou 2020, Bodeker 2024). ROBINS-I overall: Serious 5, Critical 2 (both Cameron studies), Moderate 0. The two trials with the soundest methods are the phase III colorectal trial and the placebo-controlled prostate trial. Both found no benefit.
Overall survival, intravenous stratum, by cancer type
Colorectal cancer rests on one trial, and it is a good one. Wang 2022 randomised 442 people with untreated metastatic colorectal cancer to FOLFOX with or without bevacizumab, with or without IV vitamin C at 1.5 g/kg/day: overall survival HR 1.04 (95% CI 0.81 to 1.33), median 20.7 months against 19.7. That interval rules out a large survival gain in this setting, and GRADE rates it moderate certainty.
Pancreatic cancer rests on one small trial that stopped early. Bodeker 2024 randomised 36 people with stage IV pancreatic cancer, treated 34, and stopped at an interim analysis after 26 deaths: overall survival HR 0.46 (90% CI 0.23 to 0.92), median 16.0 months against 8.3. Trials stopped early for apparent benefit tend to give inflated effect estimates, and this one is unblinded and rated High risk of bias. GRADE rates it very low certainty.
Lung cancer's one randomised result cannot be credited to vitamin C. Ou 2020 randomised 97 people with refractory advanced NSCLC to best supportive care with or without IV vitamin C at 1 g/kg, and vitamin C was always delivered together with modulated electro-hyperthermia: overall survival HR 0.3268 (95% CI 0.1582 to 0.4105) as printed. The same paper prints a progression-free survival interval whose upper bound falls below its own point estimate. GRADE rates the evidence for vitamin C itself very low, on indirectness as much as bias.
Prostate cancer's randomised result points the other way. Paller 2024, a double-blind placebo-controlled phase II trial of docetaxel with or without IV vitamin C 1 g/kg in metastatic castration-resistant prostate cancer, was suspended at a prespecified futility analysis: overall survival HR 1.98 (95% CI 0.85 to 4.58), PSA response 41% against 33% (P=0.44). GRADE rates certainty in this overall survival estimate as low, downgraded twice for imprecision: the interval runs from 0.85 to 4.58 on 47 treated patients, and overall survival was a secondary outcome in a trial powered on other endpoints.
Ovarian cancer (Ma 2014, 27 randomised, 22 in the survival set) reports survival only in graph form, with a median time to progression 8.75 months longer on ascorbate and no significance test supporting it. No hazard ratio exists to extract. Zhao 2018 (acute myeloid leukaemia, n=73) reports response and survival in a design whose randomisation is not described in the source read; it contributes no hazard ratio. Mansoor 2021 and van Gorkom 2022 report no overall survival hazard ratio.
Pooled analyses
The all-cancer IV randomised pool (4 trials, 625 randomised) gives HR 0.73 (Wald 95% CI 0.34 to 1.59; Hartung-Knapp 0.21 to 2.60), and the same result carries I-squared of 87.5%, tau-squared of 0.519 and a 95% prediction interval from 0.02 to 25.07 (Figure 1). A DerSimonian-Laird check on the same inputs gives tau-squared 0.482, close to the REML value of 0.519, so the heterogeneity finding does not rest on the estimator. A prediction interval that wide means the next similar trial could plausibly land almost anywhere. We report the pool because the protocol allowed it, and we judge it uninterpretable as a common effect.
Two sensitivity pools sharpen that judgment, and both are reported with both interval methods because at these study counts the choice of interval changes the story. Removing Ou 2020, the hyperthermia trial, moves the pooled HR to 0.99 (Wald 95% CI 0.49 to 2.01; Hartung-Knapp 0.19 to 5.26; prediction interval 0.0003 to 3591.6, which is a statement about how little three heterogeneous trials pin down). Ou's overall survival interval is also grossly asymmetric on the log scale (the distance from the estimate to the lower bound is about three times the distance to the upper bound), so its standard error is an approximation, which is a further reason to give the no-Ou pool equal weight. Pooling only the two Low risk of bias trials gives HR 1.26 (Wald 0.71 to 2.24; Hartung-Knapp 0.03 to 52.8; I-squared 51.6%, tau-squared 0.107). At two studies the Hartung-Knapp interval carries almost no information. This pool was added after risk of bias adjudication; it was not a pre-specified analysis. We demoted it at peer review from a headline to a sensitivity result: the claim it supports is the trial-level one, that neither of the two soundest trials showed a survival benefit. Its GRADE rating of Low, one level above the primary pool, comes from its two trials being at low risk of bias; the downgrades are one for imprecision (the Hartung-Knapp interval spans 0.03 to 52.8) and one for indirectness (two different cancers read as one sensitivity question).
Progression-free survival pooled across Wang, Bodeker and Paller gives HR 0.86 (Wald 95% CI 0.71 to 1.04; Hartung-Knapp 0.49 to 1.53), with I-squared 46.9% and a 95% prediction interval of 0.25 to 2.97 (Figure 2). Tau-squared was estimated at the boundary value of 0, with a Q-profile interval from 0 to 12.9, so that boundary value is not evidence of homogeneity. Wang's prespecified RAS-mutant subgroup showed a PFS HR of 0.67 (0.50 to 0.91). The trial sources read do not report an interaction test, so we record the subgroup result and make no subgroup claim.
The two non-randomised studies with adjusted hazard ratios are reported separately, and a pooled figure that appeared in the first draft was withdrawn at peer review. Simmons 2024 compares 55 allo-transplant recipients given IV vitamin C repletion with 55 propensity-matched historical controls: adjusted HR 0.50 (0.20 to 1.00). Lee 2026 follows 748 people in palliative breast care: IV vitamin C given together with mistletoe therapy as a time-dependent exposure, adjusted HR 0.561 (0.319 to 0.986). A transplant repletion cohort and a breast cohort with a second active co-therapy do not estimate a shared effect, so no pool is presented (Figure 3 shows the two estimates side by side). Both studies also carry design features that push estimates toward benefit: historical controls from an earlier transplant era (2015 to 2018 against enrolment from 2018) in one, and an eligibility rule in the other that required patients to survive and persist with treatment for at least 3 months. GRADE rates each very low certainty. The Cameron historical-control studies (mean survival above 210 days against 50 days in 1976; median 343 days against 180 days in the 1991 databank report) sit at Critical risk of bias and were not pooled.
Two included studies sit at the edge of the review's question, and their placement is a judgment a reader should see. Van Gorkom 2022 and Simmons 2024 are transplant supportive-care studies of vitamin C repletion: van Gorkom gave IV vitamin C 70 mg/kg/day during conditioning and then oral vitamin C 500 mg twice daily to day 42, reaching blood levels in the normal range (mean 69 micromol/L), not the millimolar range the antitumour hypothesis needs. They met the protocol's total-dose threshold on grams per day, so they are included, flagged, and kept out of every antitumour pool. The same intent problem applies in milder form elsewhere: Pathak 2005 tested a bundle of antioxidants, Zhao 2018 tested vitamin C as a decitabine partner on a stated TET2 mechanism in leukaemia, and Mansoor 2021 tested a four-week symptom intervention. Grouping by dose threshold is what the protocol specified; grouping by intent would be a different review, and the difference is a limitation of this one.
Oral stratum
Creagan 1979 (150 randomised, 123 evaluable) and Moertel 1985 (100 randomised) both tested oral vitamin C 10 g/day against placebo in double-blind trials and both reported no advantage in survival, progression, symptoms or response. Neither reports a hazard ratio or arm-level medians in the sources read, so the oral stratum is narrative by necessity, and the narrative is consistent: no benefit. Pathak 2005 tested a bundle of antioxidants including oral vitamin C at 6.1 g/day with chemotherapy in NSCLC and cannot isolate vitamin C.
Quality of life, symptoms and harms
Mansoor 2021 randomised 350 people with stage II to III breast cancer to standard treatment with or without IV vitamin C 25 g once weekly for 4 weeks, and reported less severe nausea, appetite loss, tumour pain, fatigue and insomnia in the vitamin C arm. Vollbracht 2011, a non-randomised breast cancer comparison, reported quality of life gains in the same direction. Bodeker 2024 reported no detriment to quality of life. Scales and constructs differ across the three, so nothing was pooled; GRADE rates a short-term symptom benefit in breast cancer as low certainty.
For grade 3 or worse adverse events, Wang 2022 counted 74 of 221 against 67 of 221 (risk ratio 1.10 from the verified counts). Bodeker 2024 reported lower rates of grade 3 and 4 haematological events with ascorbate and fewer discontinuations (0 against 5). Paller 2024 reported grade 3 to 4 adverse events of interest in 6% against 0%, on 47 treated patients. Read together, adding pharmacologic ascorbate to standard therapy did not produce a large short-term toxicity increase in these screened trial populations, and GRADE rates that narrow claim low certainty.
The named harms are a different matter. No randomised trial in the included set reported an event of haemolysis, oxalate nephropathy, kidney stones or glucose meter interference. The trials excluded people with G6PD deficiency and kidney disease, the groups in whom case reports place those harms. The trial record is silent where it is structurally unable to speak, and this review records that silence as a finding.
Sensitivity: a descriptive median-ratio comparison
Qu 2025 reported a pooled median survival ratio of 1.83 (95% CI 1.40 to 2.40) and graded it moderate certainty. On the four IV randomised trials in this review, the study-level median ratios are 1.05 (Wang), 1.93 (Bodeker), 1.68 (Ou) and 0.52 (Paller). Their unweighted geometric mean is 1.15. It carries no confidence interval and no weighting, and it is a descriptive comparison, not a pooled estimate. Excluding Ou 2020, the hyperthermia trial that the hazard-ratio sensitivity excludes, the same calculation gives 1.02. Three things changed at once against the 2025 figure: the effect measure (a ratio of medians in single trials against a pooled median ratio), the study mix (randomised trials only against a pool of eight that included historical-control and retrospective studies with large apparent effects and counted a non-randomised databank study as a randomised trial), and the trial set (neither Bodeker 2024 nor Paller 2024 was in Qu's randomised set). This review does not decompose the gap among those three changes, so no share of it is attributed to any one of them.
5 Discussion
The headline of this review is a disagreement between study designs, stated with numbers in the same place. Non-randomised and historical-control studies of high-dose vitamin C report large survival gains, with adjusted hazard ratios of 0.50 (a transplant repletion cohort) and 0.561 (a breast cohort in which vitamin C was given with mistletoe) in the two modern cohorts and far larger contrasts in the Cameron reports. That design contrast is descriptive: a formal interaction test comparing the randomised pool with the two-cohort estimates gives z 0.67, P 0.51, so no interaction is established, and with six studies in total such a test has little power either way. The test is a sensitivity calculation on the reported estimates, not a pooled effect estimate. Randomised trials report, in order of size and rigour: HR 1.04 in 442 patients (Wang, Low risk of bias; its interval runs from a 19% hazard reduction to a 33% increase, so the trial found no significant benefit rather than proving no benefit), HR 1.98 in the only double-blind placebo-controlled trial with survival follow-up (Paller, Low), HR 0.46 in a 34-patient trial stopped early (Bodeker, High), and HR 0.33 in a trial where vitamin C came packaged with heat treatment (Ou, High). A reader who weights evidence by design quality reaches the opposite conclusion from a reader who pools everything. One asymmetry in our own reading deserves naming: we are sceptical of Bodeker's early stop for benefit and we report Paller's stop for futility without the same scepticism. The two stops are not symmetric in effect, since stopping for benefit inflates an estimate while a prespecified futility stop mainly truncates one, but a reader is entitled to see both named.
That pattern has a history in this exact field. Cameron's historical-control results in the 1970s were large, and the blinded Mayo trials that followed found nothing. The lesson the field drew then was about control groups, and the same lesson is available now, in hazard ratios.
Where does that leave the pancreatic signal? Bodeker 2024 is the one result that could change practice if it holds: a doubling of median survival, 16.0 months against 8.3, in a cancer where gains are usually counted in weeks. It is also the kind of result (34 treated patients, stopped at interim, unblinded, 90% intervals) of the kind that tends to shrink, and sometimes vanish, in a confirmatory trial. The honest statement is narrow: one small trial found a large effect, the trial's own authors present it as design data for a phase III study, and no phase III test of it exists yet. Our search found no completed phase III trial of pharmacologic ascorbate in pancreatic or colorectal cancer beyond VITALITY, whose overall survival result was null.
Limitations of this review are specific. Five databases with no access in this environment (Embase, Scopus, Web of Science, CENTRAL, CINAHL) were not searched, the WHO ICTRP portal could not be searched by API, and OpenAlex was blocked by a rate limit, and a review that claims completeness should say so plainly; the plain language summary wording has been narrowed accordingly; the two studies found only by citation chasing (a 1991 databank report and a 2018 leukaemia trial) show the database net had holes, although neither changes any pooled result. Two records (R1550, R2607) could not be classified from their titles alone and are flagged for the author rather than silently dropped. Moertel 1985 and Creagan 1979 were extracted at abstract level for several fields because the full texts sit behind paywalls; the fields affected are listed in the extraction notes, and no pooled number depends on them. Zhao 2018 is described throughout as a comparative study whose randomisation is not described in its report; it is assessed with RoB 2 for transparency, flagged wherever the counts are given, and it contributes no estimate to any pool. The Bodeker intervals are 90% intervals, handled as such in the standard error calculation. One included study (Lee 2026) is dated 2026 in its journal's online-first sequence and was retrieved in the November 2026 search.
Strengths are also specific. Every number in the abstract traces to the settled extraction sheet or the analysis output. Screening, extraction and risk of bias were each done twice, independently, with disagreements settled by reading rather than by rule, and the discrepancy rates are reported as results (Sections: Methods; team log in the supplement). The competing 2025 review's own included studies were cross-checked one by one, which is how its two missed-by-our-search studies and its mislabelled databank trial came to light.
6 Conclusions
Implications for practice. This review makes no recommendation that any person should or should not receive vitamin C, at any dose, by any route. What the evidence establishes is narrower. Randomised trials do not show a survival benefit for high-dose vitamin C added to standard cancer therapy, and the two most rigorous trials are the two that found none (per-trial HRs 1.04 and 1.98; their sensitivity pool is HR 1.26 with intervals too wide at two studies to stand alone). For harms, the same trials show no large short-term toxicity increase in screened populations, and they are silent on the named specific harms for structural reasons given in the Results. The claim of moderate-certainty benefit published in 2025 does not survive a randomised-only, hazard-ratio analysis (four-trial pool HR 0.73 with a prediction interval of 0.02 to 25.07, and 0.99 (Hartung-Knapp 0.19 to 5.26) once the hyperthermia co-intervention trial is removed). High-dose IV vitamin C for cancer is investigational. The FDA has not approved it as a treatment for cancer, and the NCI and Memorial Sloan Kettering both classify it for trial settings.
Implications for research. The trial that would settle the pancreatic question is specifiable: multicentre, double-blind where feasible, powered on overall survival, with plasma ascorbate levels measured and reported, and no interim stop for benefit short of a stringent boundary. For colorectal cancer, VITALITY's RAS-mutant subgroup result (a progression-free survival subgroup in a trial whose overall survival result was null, with no interaction test reported) justifies a trial with a pre-specified interaction test, not a claim and not a design that assumes the subgroup effect. In pancreatic cancer, KRAS mutation is close to universal, so stratification by KRAS status is not a workable design there; the workable stratifiers are measured plasma ascorbate exposure and other candidate markers. On plasma levels, the trials that reported them give concrete targets to copy: Ma 2014 titrated to a peak of 350 to 400 mg/dL (20 to 23 mM) measured at end of infusion, and the glioblastoma programme found 87.5 g three times weekly consistently reached at least 20 mM. Future trials should report hazard ratios with 95% intervals for survival outcomes as a matter of course; two otherwise eligible trials in this review could contribute nothing to any pool because they reported survival in graphs or in words. Harms reporting should name the specific events of interest (haemolysis, oxalate nephropathy, stones, glucose meter interference) even when the count is zero, and should state who was screened out.
Declarations
Correspondence. Raphael Malikian, Palmdale, CA, USA. Email rtmalikian@gmail.com. ORCID 0009-0008-4326-4205.
Funding. None. This review received no funding from any source.
Conflicts of interest. None. The author is fully independent, with no industry affiliations of any kind.
Author contributions. Raphael Malikian is the sole accountable author. He set the question and the protocol, reviewed the flagged records list, and takes responsibility for the final text.
AI assistance. This review was carried out by a team of AI agents under the author's direction, following the published Robust Standard v2 (dual screening, dual extraction, adjudication by reading, duplicate risk of bias). AI tools were used for searching, screening, extraction, statistical analysis, drafting and editing. The draft received structured multi-agent pre-submission peer review under the author's published protocol. The human author reviewed every comment and is accountable for the final text. This is pre-submission review by AI agents. It is not journal peer review and is not described as such anywhere in the submission package.
Ethics. No human participants, no patient-level data. All data come from published reports.
Data availability. Search log and strings, deduplicated record set, both screening passes, adjudication files, both extraction passes, settlement record, risk of bias files, analysis code (Python), analysis-ready datasets (CSV) and figure files are held in the review workspace and prepared for deposit with the GitHub and Zenodo packages that accompany this manuscript. Nothing has been deposited or submitted at the time of writing.
References
- Cameron E, Pauling L. Supplemental ascorbate in the supportive treatment of cancer: prolongation of survival times in terminal human cancer. Proc Natl Acad Sci USA. 1976;73(10):3685-3689. doi:10.1073/pnas.73.10.3685
- Cameron E, Campbell A. Innovation vs. quality control: an 'unpublishable' clinical trial of supplemental ascorbate in incurable cancer. Med Hypotheses. 1991. doi:10.1016/0306-9877(91)90127-k. PMID 1787807
- Creagan ET, Moertel CG, O'Fallon JR, et al. Failure of high-dose vitamin C (ascorbic acid) therapy to benefit patients with advanced cancer: a controlled trial. N Engl J Med. 1979;301(13):687-690. doi:10.1056/NEJM197909273011303
- Moertel CG, Fleming TR, Creagan ET, et al. High-dose vitamin C versus placebo in the treatment of patients with advanced cancer who have had no prior chemotherapy: a randomized double-blind comparison. N Engl J Med. 1985;312(3):137-141. doi:10.1056/NEJM198501173120301
- Pathak AK, et al. Chemotherapy alone vs chemotherapy plus high dose multiple antioxidants in patients with advanced non small cell lung cancer. J Am Coll Nutr. 2005;24(1):16-21. doi:10.1080/07315724.2005.10719438
- Vollbracht C, Schneider B, Leendert V, Weiss G, Beuth J. Intravenous vitamin C administration improves quality of life in breast cancer patients during chemo-/radiotherapy and aftercare: results of a retrospective, multicentre, epidemiological cohort study in Germany. In Vivo. 2011;25(6):983-990. PMID 22021693
- Ma Y, Chapman J, Levine M, Polireddy K, Drisko J, Chen Q. High-dose parenteral ascorbate enhanced chemosensitivity of ovarian cancer and reduced toxicity of chemotherapy. Sci Transl Med. 2014;6(222):222ra18. doi:10.1126/scitranslmed.3007154
- Zhao H, et al. The synergy of vitamin C with decitabine activates TET2 in leukemic cells and significantly improves overall survival in elderly patients with acute myeloid leukemia. Leuk Res. 2018. doi:10.1016/j.leukres.2017.12.009. PMID 29331774
- Ou J, Zhu X, Chen P, et al. A randomized phase II trial of best supportive care with or without hyperthermia and vitamin C for heavily pretreated patients with non-small cell lung cancer. J Adv Res. 2020;24:175-182. doi:10.1016/j.jare.2020.03.004
- Ou J, et al. A retrospective study of gemcitabine and carboplatin with or without intravenous vitamin C on patients with advanced triple-negative breast cancer. Integr Cancer Ther. 2020. doi:10.1177/1534735419895591. PMID 32070148
- Mansoor F, et al. Impact of intravenous vitamin C administration in reducing severity of symptoms in breast cancer patients during treatment. Cureus. 2021. doi:10.7759/cureus.14867. PMID 34113504
- Wang F, He MM, Xiao J, et al. A randomized, open-label, multicenter, phase 3 study of high-dose vitamin C plus FOLFOX with or without bevacizumab versus FOLFOX with or without bevacizumab in unresectable untreated metastatic colorectal cancer (VITALITY). Clin Cancer Res. 2022;28(19):4232-4239. doi:10.1158/1078-0432.CCR-22-0655
- van Gorkom GNY, et al. Vitamin C supplementation in autologous stem cell transplantation. Nutrients. 2022. doi:10.3390/nu14224784
- Ou J, et al. SERPINE1 and SERPINB7 as potential biomarkers for intravenous vitamin C treatment in non-small-cell lung cancer. Free Radic Biol Med. 2023. doi:10.1016/j.freeradbiomed.2023.10.391. PMID 37838303
- Bodeker KL, Smith BJ, Berg DJ, et al. A randomized trial of pharmacological ascorbate, gemcitabine, and nab-paclitaxel for metastatic pancreatic cancer. Redox Biol. 2024;77:103375. doi:10.1016/j.redox.2024.103375
- Paller CJ, Zahurak ML, Mandl A, et al. High-dose intravenous vitamin C combined with docetaxel in men with metastatic castration-resistant prostate cancer: a randomized placebo-controlled phase 2 trial. Cancer Res Commun. 2024;4(8):2174-2182. doi:10.1158/2767-9764.CRC-24-0225
- Simmons G, et al. Intravenous vitamin C supplementation in allogeneic hematopoietic cell transplant recipients. eJHaem. 2024. doi:10.1002/jha2.995. PMID 39415933
- Lee JY, et al. High-dose intravenous vitamin C and mistletoe therapy in breast cancer: a prospective cohort study on overall survival. Cancers. 2026. doi:10.3390/cancers18193094
- Qu J, et al. Overall and progression-free survival of patients with malignant neoplasm following intravenous vitamin C: a systematic review and meta-analysis. Int J Vitam Nutr Res. 2025;95(3). doi:10.31083/IJVNR37372. PMID 40613397
- Shahzad et al. Vitamin C as an adjunctive therapy in cancer: a systematic review and meta-analysis of outcomes in solid and hematologic malignancies. Am J Clin Oncol. 2026. doi:10.1097/COC.0000000000001353. PMID 42507830
- Alangari A, Arif J, Al Qureshah F, Alkhodairy F. Clinical benefits and risks of high-dose intravenous vitamin C: a systematic review. J Med Life. 2026;19(1):16-23. doi:10.25122/jml-2025-0176. PMID 41815850
- National Cancer Institute. Intravenous Vitamin C (PDQ): Health Professional Version. Version dated 13 May 2025. cancer.gov.
- Nielsen TK, Højgaard M, Andersen JT, et al. Weekly ascorbic acid infusion in castration-resistant prostate cancer patients: a single-arm phase II study. Transl Androl Urol. 2017;6(3):517-528. doi:10.21037/tau.2017.04.42 [Signals table only.]
- Allen BG, Bodeker KL, Smith MC, et al. First-in-human phase I clinical trial of pharmacologic ascorbate in previously treated or newly diagnosed solid tumors / glioblastoma phase I report. Clin Cancer Res. 2019;25(22):6590-6597. doi:10.1158/1078-0432.CCR-19-0594 [Signals table only.]
- Petronek MS, Monga V, Bodeker KL, et al. Single-arm phase II of pharmacological ascorbate in glioblastoma. Clin Cancer Res. 2024;30(2):283-293. doi:10.1158/1078-0432.CCR-22-3952 [Signals table only.]
Figures
Figure 1. Overall survival, intravenous ascorbate randomised trials, with the REML pooled diamond and the 95% prediction interval (bar below the diamond). File: figures/fig1_os_iv_rct.png Figure 2. Progression-free survival, intravenous ascorbate randomised trials (Ou 2020 not pooled; its printed interval is invalid). File: figures/fig2_pfs_iv_rct.png Figure 3. Overall survival, the two non-randomised studies with adjusted hazard ratios, shown side by side and not pooled. File: figures/fig3_os_cohorts.png
Supplement (held with the review, prepared for deposit)
PRISMA flow data; the cross-check of Qu 2025's eight included studies against this review's classification; full search strategies and log; full-text exclusion list with reasons; unclassified records flagged for the author (R1550, R2607); both extraction passes and the settlement record; risk of bias tables; LLM team log; PRISMA 2020 and PRISMA-S checklists.