Innovation & Tech

Fujitsu AI flags 51 lung-cancer trial candidates from 22,334 records

Fujitsu Japan, the National Cancer Center Hospital East and Paradigm Health used an LLM to screen 22,334 patient records, with physicians reviewing 51 promising trial candidates out of 57 flagged.

By Daniel Okafor

3 min read

Fujitsu and Paradigm test AI screening for Japan trials
Fujitsu and Paradigm test AI screening for Japan trialsAI-generated

What's News

  • Retrospective study screened 22,334 records from the National Cancer Center Hospital East's respiratory medicine department, June through end of October
  • Fujitsu processed 3,027 lung-cancer patients; 540 records with EGFR mutations were checked against five eligibility criteria
  • Paradigm's platform selected 57 priority candidates and physicians judged 51 to be promising
  • Overnight automated processing cut the number of records requiring manual review by approximately 89 percent
  • Fujitsu targets a commercial screening-to-enrollment service by the end of fiscal 2027

Fujitsu Japan, the National Cancer Center Hospital East and US-based Paradigm Health used a large language model to screen 22,334 patient records and flagged 51 promising clinical-trial candidates, according to a Fujitsu Japan statement released Thursday. The retrospective verification study asked a narrow question: can automated pre-screening find eligible patients faster than manual chart searches?

How did the screening pipeline work?

The system does not decide eligibility. It does two jobs instead. It pulls patient information out of electronic medical records and structures it into fields a computer can query, using a large language model. It then feeds that structured data into Paradigm's clinical trial matching platform, which ranks records against a trial's criteria and attaches clinical evidence to each selection.

The chain matters because each stage carries a different risk. The model can misread or flatten a record. The platform can rank a chart on thin evidence. Physician review is what catches both errors before a name reaches a trial site.

The study covered records from the hospital's respiratory medicine department, collected between June and the end of October. For a simulated trial built around EGFR gene mutations, Fujitsu processed data from 3,027 lung-cancer patients.

The funnel, in numbers:

  • 22,334 patient records from the respiratory medicine department, June through October
  • 3,027 lung-cancer patients whose data Fujitsu processed
  • 540 records carrying EGFR mutations, checked against five eligibility criteria
  • 57 priority candidates selected by Paradigm's platform, with clinical evidence attached
  • 51 records that physicians determined were promising candidates
  • Approximately 89 percent fewer records requiring manual review

Paradigm's platform selected the 57 and presented the supporting evidence. Physicians then reviewed the records and determined that 51 were promising candidates, Fujitsu said. No patient was enrolled. Fujitsu stressed that the output “did not determine whether patients are eligible for a real trial.” Final eligibility remains subject to trial requirements, physician review and patient consent.

How much manual review did the system remove?

Automated overnight processing reduced the number of records requiring manual review by approximately 89 percent, according to the study. Roughly nine in every ten records avoided a human read, and that ratio is the number hospital staffing models care about.

The reason is mundane. Treatment histories, test results and genomic information sit in scattered places, and staff otherwise apply standardized criteria by hand. Fujitsu said the approach could help medical staff apply those criteria without manually searching dispersed records. In a department holding tens of thousands of patient files, that difference separates reading every chart from reading a shortlist.

What does “drug loss” mean?

Fujitsu has framed the project as part of efforts to reduce “drug loss,” a term the company uses for treatments available overseas that are not developed or approved in Japan. The recruitment problem sits alongside that gap. As trials specify narrower genetic and biomarker criteria, clinical staff find it harder to identify eligible patients inside their own files. Pre-screening targets that bottleneck.

The two problems are related but not interchangeable, because a therapy Japan has not approved cannot be trialled at home, and a trial nobody recruits for yields no data to support approval in the first place.

What does Fujitsu plan to sell, and when?

Fujitsu plans to commercialize a service covering the path from candidate screening through clinical trial enrollment by the end of fiscal 2027. The company said the partners intend to promote adoption by medical institutions across Japan. Paradigm had already reported preliminary findings at the European Society for Medical Oncology Congress in October. The November statement adds the detailed workflow, physician-review results and commercialization target.

Fujitsu disclosed no price, no named hospital customers and no timeline for a live trial outside the simulation. Those gaps will matter more than the funnel from 22,334 records down to 51 physician-approved names, because commercialization by fiscal 2027 will be judged on one measure: whether hospitals beyond the National Cancer Center Hospital East use the service to assemble real candidate lists for real studies.

Source: technode.global

Share this article:

More from Daniel Okafor

Daniel Okafor

Show full bio

Correspondent covering money, funding and market shifts around entrepreneurship, startups and venture funding at Business Bearings.

2,443 articles

Related articles

« Previous article