Researchers in Japan have shown that survival outcomes for a modern lung cancer immunotherapy can be reconstructed from the free-text pages of electronic medical records held under a national data-sharing law, producing results that line up closely with a landmark randomized trial. The historical cohort study, published in Drugs – Real World Outcomes, analyzed 75 patients with non-small cell lung cancer treated with atezolizumab and found a median progression-free survival of 3.5 months, a figure the authors describe as comparable to the 4.2 months reported for the Japanese subgroup of the international OAK trial. The work is less a clinical discovery than a proof of concept for a regulatory question: can pseudonymized real-world data collected under Japan’s Next-Generation Medical Infrastructure Act reproduce the kind of evidence normally generated only inside randomized controlled trials?

The study was conducted as a historical cohort analysis using the Millennium Medical Record database, maintained by the Life Data Initiative, a certified pseudonymized medical data handling operator under the Act. The team drew on medical information collected between 1 October 2015 and 31 October 2022, while the study itself ran from September 2023 to February 2024. The benchmark was the OAK trial, a randomized, controlled, open-label, international phase 3 study that compared atezolizumab monotherapy with docetaxel in previously treated patients with locally advanced or metastatic disease, with overall survival as its primary endpoint and investigator-assessed progression-free survival among its secondary endpoints.

Building the cohort required a funnel of increasingly strict filters. The Millennium Medical Record database held 1,651,706 cases in total. The population was first narrowed to 745,777 cases from 12 medical institutions where all three data sources used in the study—medical claims, Diagnosis Procedure Combination survey data, and electronic medical records—were available and could be linked at the patient level. Selecting patients who had received atezolizumab monotherapy with no prior prescription history of small cell lung cancer treatments—irinotecan, etoposide, nogitecan, and amrubicin—reduced the pool to 100 cases. After excluding records with missing chemotherapy documentation or inconsistencies in diagnosis dates, 75 patients formed the primary analysis set.

The most technically demanding step was extracting outcomes from unstructured text. Progression-free survival was defined as the time to progressive disease according to RECIST version 1.1, or death from any cause, while overall survival was measured as time to death from any cause. The event date for overall survival came from a structured field, the date and time of death column in the electronic medical record. For progression, however, the event date was determined from the entry date in the medical record where progressive disease or exacerbation was explicitly documented in the text. Trained abstractors performed manual review of the records to identify these entries, a labor-intensive process the authors argue ensured the robustness of the assessments. Survival curves were analyzed with the Kaplan–Meier method and the log-rank test, and a Cox proportional hazard model estimated hazard ratios with 95% confidence intervals, using Python 3.8.5 and the lifelines 0.27.8 survival analysis library. Atezolizumab initiation was defined as day 0, and patients lost to follow-up, for example through transfer to another facility, were censored.

The headline result was the progression-free survival figure. Median PFS in the real-world cohort was 3.5 months, with a 95% confidence interval of 2.3 to 8.9 months. In the OAK trial’s Japanese subgroup, median PFS was 4.2 months, with a 95% confidence interval of 2.8 to 4.4 months. The authors present this as evidence that survival time analysis and curve plotting are feasible with data drawn from unstructured text under the legal framework, supporting the applicability of real-world data for estimating survival functions. Subgroup analyses reinforced internal validity in a qualitative sense. Median PFS was 155 days in patients with stage 2 or lower disease compared with 91 days in those with stage 3 or higher, yielding a hazard ratio of 1.43 with a 95% confidence interval of 0.56 to 3.62 and a p-value of 0.45. Patients with an ECOG performance status of 1 or better had a median PFS of 122 days versus 54 days for those with a status of 2 or worse, a hazard ratio of 1.82 with a confidence interval of 0.80 to 4.14 and a p-value of 0.15. Neither difference reached statistical significance, but the authors note that the numerical trends toward shorter PFS in advanced stage and poorer performance status were consistent with established clinical knowledge.

Overall survival told a more complicated story. The median was not reached, because the survival rate remained above 50% throughout the observation period, and the 1-year survival rate was 83.0%. The authors attribute this partly to censoring: the Millennium Medical Record system cannot track deaths that occur after a patient transfers to another institution, so patients who leave are recorded as lost to follow-up rather than as deaths. For context, the OAK trial reported a median overall survival of 21.3 months, with a 95% confidence interval of 11.0 months to not estimable, for its 36 Japanese patients. The authors also flag a more serious concern in their limitations section: overall survival in this study may be overestimated precisely because of those tracking limitations, a bias that any regulatory use of such data would need to address.

The legal framework underlying the study is central to its significance. Japan’s Next-Generation Medical Infrastructure Act, enacted in 2017 and in force since 2018, promotes the secure use of medical data for research, innovation, and healthcare improvement. Hospitals and medical institutions may provide patient information to government-certified operators, who process it into anonymized or pseudonymized forms in which individuals cannot be readily identified, and approved users including researchers, universities, pharmaceutical companies, and medical device developers can then access it. Certified organizations must meet strict requirements on cybersecurity, governance, and compliance, and patients are informed that their data may be used and generally hold an opt-out right to request exclusion.

That linkage is where the authors see the path forward for the study’s weakest outcome. Because the National Database captures death notifications, future integration is expected to improve overall survival accuracy by incorporating post-transfer mortality data, addressing the censoring that left the median unreached. Achieving complete outcome tracking through such linkage is described as a critical step toward broader regulatory applications. With improved accuracy in survival outcomes, the authors argue, this infrastructure could eventually support the generation of external control arms for single-arm trials, particularly for rare cancers or specific molecular subgroups where randomized controlled trials are difficult to conduct, potentially accelerating drug development and approval processes in Japan. The study also clarified which effectiveness metrics are currently assessable and which are not, emphasizing the importance of prior validation before implementing artificial intelligence-based systems on such data.

The findings should be read with appropriate caution. The primary analysis set of 75 patients is small, a consequence of strictly applying selection criteria to isolate patients receiving atezolizumab monotherapy within a legally regulated database, which reduced statistical power for the subgroup analyses. Although the cohort is larger than the 36-patient Japanese subgroup of the OAK trial, the authors state that the findings should be interpreted with caution and that larger multi-database studies are warranted to confirm the results. They also acknowledge the inherent limitations of observational studies, including potential selection bias and immortal time bias, while arguing that the methodology shows potential for broader applications in real-world data analysis. The study was approved by the institutional review board at each participating facility, data extraction and anonymization followed institutional approval, and the anonymized dataset was stored on secure computing servers with cybersecurity protocols, with statistical analyses conducted in a dedicated environment with controlled access. As a demonstration that unstructured medical record text can yield trial-comparable progression-free survival under Japan’s regulated data framework, the study marks a measured but meaningful step toward real-world evidence that regulators may one day trust.

Subject of Research: Reconstructing survival outcomes for atezolizumab-treated lung cancer patients from unstructured electronic medical records under Japan's Next-Generation Medical Infrastructure Act

Article Title: Feasibility of Reconstructing Survival Outcomes from Unstructured Electronic Medical Records under Japan’s Next-Generation Medical Infrastructure Act: A Historical Cohort Study of Atezolizumab in Non-small Cell Lung Cancer

Article References: Nakatsui, M., Nishida, Y., Nozue, S., Sakata, Y., Azuma, M., Abe, R., Okamoto, R., Nakagami, Y., & Okuno, Y. (2026). Feasibility of Reconstructing Survival Outcomes from Unstructured Electronic Medical Records under Japan’s Next-Generation Medical Infrastructure Act: A Historical Cohort Study of Atezolizumab in Non-small Cell Lung Cancer. Drugs – Real World Outcomes, 13(3), 421-429. https://doi.org/10.1007/s40801-026-00570-w

Image Credits: AI Generated

DOI: 10.1007/s40801-026-00570-w

Keywords: real-world evidence, electronic medical records, Next-Generation Medical Infrastructure Act, atezolizumab, non-small cell lung cancer, progression-free survival, overall survival, OAK trial, external control arms, historical cohort study, Japan, pseudonymized data