An arthritis drug helped patients with severe COVID-19

AIBD-0042-benefit-mortality, AIBD-0042-benefit-recovery

Improved Physical Health [1] [2] [5]

Patient benefits were observed. Hospital savings were modeled. WHO recommended wider treatment. These studies did not measure population-wide effects. [1] [2] [5] [7] [10]

People, organizations, and society. [1] [2] [5] [7] [10]

Hospitalized patients with COVID-19 and the hospitals and health systems treating them. [1] [2] [5] [10]

The trials enrolled 10,714 participants in total, including controls. That total does not identify how many people benefited. The publications give no worldwide count. [1] [2] [5]

ACTT-2 patients on high-flow oxygen or noninvasive ventilation recovered in a median of 10 rather than 18 days. This subgroup estimate is less precise. RECOVERY found no clear differences among subgroups. [1] [2] [5] [10]

The trials followed patients for 28 days. COV-BARRIER investigators also examined deaths through day 60 in an exploratory analysis. The economists projected costs and health over a lifetime. [1] [2] [5] [10]

Researchers deliberately sought a treatment, and investigators planned to measure mortality. [1] [2] [4]

Serious adverse events affected 15% of baricitinib recipients and 18% of controls in COV-BARRIER. RECOVERY investigators reported 12 suspected serious reactions. Short follow-up leaves rare and long-term harms uncertain. [1] [2] [5] [10]

Benevolent Platform [3] [4] [8]

SVO, a rule-based model that extracts relationships from sentences. [4]

The configuration used in January 2020. [3] [4]

Editor’s note. Smith and colleagues describe the January 2020 work but give no release identifier for SVO (reference 4, p. 4).

BenevolentAI [3] [4]

The system extracted biomedical relationships, searched the graph, and analyzed networks of proteins. [4]

Researchers added relationships from coronavirus papers to the knowledge graph. They used the Louvain algorithm to group connected proteins and examined the networks with interactive tools. [3] [4]

The system drew from biomedical literature and databases to produce relationships and protein networks that researchers used to assess candidate drugs. [3] [4]

Scientists directed the analysis and selected the drug. Clinicians controlled treatment. [3] [4]

Scientists searched the graph, examined protein networks, and assessed medicines. Clinical investigators then tested baricitinib. [1] [2] [3] [4] [5]

Researchers chose the candidate drug. Clinicians prescribed treatment under trial protocols, with oversight from ethics committees and monitors. [1] [2] [4] [5]

Researchers used AI to investigate a new use for an existing medicine. [3] [4]

Researchers used the system to find relationships among proteins, diseases, and medicines before selecting baricitinib. [1] [2] [3] [4] [5]

ACTT-2, COV-BARRIER, RECOVERY, and Ohsfeldt and colleagues’ 2021 economic study. [1] [2] [5] [10]

Investigators randomly assigned treatment in all three trials. ACTT-2 and COV-BARRIER concealed assignments, while RECOVERY disclosed them. The economic study simulated costs and health. [1] [2] [5] [10]

Investigators used clinical assessments, follow-up forms, and health records. The economists combined trial findings with treatment prices and costs of care. [1] [2] [5] [10]

The trials enrolled hospitalized patients in 2020–2021: 1,033 in ACTT-2, 1,525 in COV-BARRIER, and 8,156 in RECOVERY. The economic study modeled insured U.S. adults. [1] [2] [5] [10]

The primary outcomes were recovery time in ACTT-2, disease progression in COV-BARRIER, and deaths within 28 days in RECOVERY. The economists estimated costs and quality-adjusted life-years. [1] [2] [5] [10]

Patients received baricitinib or no baricitinib alongside the same background care. ACTT-2 gave remdesivir to both groups. [1] [2] [5] [10]

Median recovery was 7 rather than 8 days in ACTT-2. Deaths by day 28 were 8% versus 13% in COV-BARRIER and 12% versus 14% in RECOVERY. [1] [2] [5] [10]

The software probably reduced the work needed to assemble and examine relevant biomedical findings.

Editor’s note. Smith and colleagues report about 40,000 added relationships and a starting network of 556 proteins (reference 4). These quantities support an inference of substantial assistance, but the authors did not time a comparable search without AI.

COV-BARRIER investigators did not establish less disease progression, which prevented confirmatory testing of mortality under their analysis plan. RECOVERY participants knew their treatment. Lilly funded COV-BARRIER and the economic study. None compared discovery with and without AI. [1] [2] [5] [10]

ACTT-2 assigned 515 patients to baricitinib and 518 to control. COV-BARRIER assigned 764 and 761. RECOVERY assigned 4,148 and 4,008. [1] [2] [5] [10]

Trial registrations, protocols, analysis plans, and appendices are public. The complete historical AI system and economic model are unavailable, and some cost inputs are unpublished. [1] [2] [4] [5] [10]

Estimated hospital savings and the lifetime cost of additional health from treatment with baricitinib. The study assigns no separate monetary value to AI. [10]

The model estimates US$1,932 less in net hospital costs per patient. Over a lifetime, costs rise US$17,276 and health improves by 0.6703 quality-adjusted life-years. [10]

U.S. dollars, with inputs from 2020–2021. Some costs were adjusted to 2020 prices, but the inputs do not share one price year. [10]

Hospital costs cover one admission per patient. Costs to insurers cover each patient’s remaining lifetime. [10]

The amounts come from an economic model, without observed cash savings. [10]

Patients gain health. Hospitals pay for treatment and receive reimbursement from Medicare, Medicaid, and private insurers. [10]

Hospital costs are net of reimbursement. Lifetime costs include healthcare and missed work. The model excludes long COVID, COVID-related readmissions, and the value of freed beds. [10]

The model includes medicines, hospital care, later healthcare, and missed work. It excludes the costs of developing and operating the AI system. [10]

The researchers compared baricitinib plus standard care with placebo plus the same care. Assuming only lower mortality reverses hospital savings to US$1,978 in additional costs per patient. [10]

The researchers combined trial results with U.S. costs and applied a 3% annual discount rate. Hospital savings depend on clinical differences that were not all statistically conclusive. [10]

At US$50,000 per quality-adjusted life-year, AIBD values the additional health at US$16,239 per patient after costs: (50,000 × 0.6703) − 17,276. Cash returns and payment dates were not reported.

BenevolentAI researchers selected the drug. U.S. government researchers, Lilly, and Oxford organized trials. The FDA authorized treatment, and WHO recommended it. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

BenevolentAI operated in the United Kingdom. Trials ran in Argentina, Brazil, Denmark, Germany, India, Italy, Japan, Mexico, Russia, Singapore, South Korea, Spain, the United Kingdom, and the United States. [1] [2] [3] [4] [5] [9]

Kenneth Mulvany controlled BenevolentAI at the end of 2019 and held 43.18% of its general voting shares in October.

Editor’s note. The percentage is calculated from his 881,000 shares and the 2,040,452 voting shares in the 15 October 2019 register (reference 12). The December allotment left that total unchanged, but individual holdings could have changed before the January research.

The researchers used BenevolentAI’s graph, SciPy, and 11 biomedical resources, including KEGG, Reactome, and SIGNOR. Their methods paper does not identify the hardware or hosting provider. [4]

Singapore’s Temasek invested in BenevolentAI. The U.K. researchers used Japanese and Italian scientific resources before trials in several countries. [1] [3] [4] [5]

Editor’s note. The geographic links follow the producers of KEGG and SIGNOR, which Smith and colleagues name in reference 4. They identify international knowledge reuse, not the route of a data transfer. Temasek funded the company as a whole.

This work may strengthen the United Kingdom’s ability to find treatments during a new outbreak.

Editor’s note. The inference rests on the U.K. team’s January 2020 proposal and the subsequent international trials (references 1–5). Those studies establish a useful treatment, but give no estimate of how much earlier AI made it available.

Testing the treatment again requires comparable patients and background care. Repeating the discovery requires the original graph, software configuration, and researchers’ decisions. [1] [2] [4] [5]

COV-BARRIER and RECOVERY both found fewer deaths with baricitinib. The trials tested the medicine, without repeating the AI discovery process. [1] [2] [5]

Oxford’s RECOVERY investigators independently confirmed lower mortality in July 2022. [2] [6] [7]

BenevolentAI used its internal platform. Its methods are published, but the complete January 2020 system is unavailable. Baricitinib is a prescription medicine. [3] [4] [6] [7] [8]

The FDA considered emergency authorization. WHO considered treatment guidance for severe or critical COVID-19. [6] [7]

The FDA relied on ACTT-2. WHO weighed the clinical evidence, harms, affordability, and access. [6] [7]

The FDA authorized baricitinib with remdesivir on 19 November 2020. WHO recommended it with corticosteroids on 14 January 2022. [2] [6] [7]

Both agencies sought better survival or recovery with acceptable harms. They did not quantify the additional use or health improvement that their decisions caused. [1] [2] [5] [6] [7]

AIBD-0042-source-0001 through AIBD-0042-source-0010, plus eight supplementary references. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

References 1–10 cover discovery, trials, and costs. References 11–18 cover ownership, funding, and geography. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

English. No translation was needed. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

The developers supplied four sources. Four were independent, and two involved the drug sponsor. Lilly funded COV-BARRIER and the economic study. Manufacturers supplied ACTT-2 medicines and protocol advice, without funding the trial. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

The FDA and WHO based their conclusions on clinical trials. BenevolentAI’s announcements repeat its discovery account and the trial results. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

The discovery letter, COV-BARRIER figure, and RECOVERY appendix have corrections. Sorrell’s supplementary table and the January 2022 WHO guideline PDF were inaccessible, although the article and WHO announcement were available. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

Researchers used AI to select baricitinib for testing. Clinical investigators subsequently found that patients recovered faster and fewer died. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

References 3–4 describe the discovery, 5 reports recovery, 1–2 report mortality, 6–7 record agency decisions, and 10 estimates costs. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

Randomized trials establish the treatment benefit. The developers describe the AI workflow. No trial compares discovery with and without AI. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

The publications omit the exact software release, a comparison without AI, observed savings, and the number of people helped worldwide. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

Hospitalized patients, hospitals, and the wider population that could receive treatment. [1] [2] [5] [7] [10]

For patients: 5 for fewer deaths and 2 for faster recovery. The broader assessments remain 2 for organizations and 4 for society. [1] [2] [5] [7] [10]

Lower mortality preserves lives. Faster recovery is a measurable health improvement. Shorter hospital stays could free capacity, and wider access could extend the treatment’s benefits. [1] [2] [5] [7] [10]

Trials established a benefit for patients. Hospital savings were modeled, and wider health benefits remain intended. [1] [2] [5] [7] [10]

AIBD’s five-level scale, applied on 25 October 2026.

AIBD, assisted by Codex, on 25 October 2026.

The hospital and societal grades describe possible benefits. A rule requiring observed benefits would leave those two grades unassigned. [1] [2] [5] [7] [10]

The solid fill denotes independently studied treatment benefits. It does not mean that the projected hospital and societal benefits occurred. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

Repurposing a medicine for COVID-19. [1] [2] [3] [4] [5]

ACTT-2, COV-BARRIER, and RECOVERY tested baricitinib after the initial proposal.

Biomedicine and hospital treatment of infectious disease. [1] [2] [5]

Researchers sought existing medicines that could treat COVID-19. [3] [4]

Researchers identified baricitinib in January 2020 and reported it on 4 February 2020. COV-BARRIER mortality results appeared on 1 September 2021. [1] [2] [3] [4] [5]

Researchers used the AI system in January 2020. Clinical trials and treatment authorization followed. [2] [3] [4] [6] [7]

Revision a8e1b005, reviewed on 25 October 2026.

ACTT-2: NCT04401579. COV-BARRIER: NCT04421027. RECOVERY: NCT04381936 and ISRCTN50189673. [1] [2] [5]

The benefit matches Trustible’s “Improved Physical Health” category, published on 1 May 2024.

Researchers used AI to identify baricitinib for COVID-19 trials. Clinical investigators later found that patients recovered sooner and fewer died. [1] [2] [3] [4] [5]

Ownership, funding, and economic estimates were added on 25 October 2026. Earlier versions remain available. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10]