Although machine learning (ML) has shown promise across disciplines, out-of-sample generalizability is concerning. This is currently addressed by sharing multi-site data, but such centralization is challenging/infeasible to scale due to various limitations. Federated ML (FL) provides an alternative paradigm for accurate and generalizable ML, by only sharing numerical model updates. Here we present the largest FL study to-date, involving data from 71 sites across 6 continents, to generate an automatic tumor boundary detector for the rare disease of glioblastoma, reporting the largest such dataset in the literature (n = 6, 314). We demonstrate a 33% delineation improvement for the surgically targetable tumor, and 23% for the complete tumor extent, over a publicly trained model. We anticipate our study to: 1) enable more healthcare studies informed by large diverse data, ensuring meaningful results for rare diseases and underrepresented populations, 2) facilitate further analyses for glioblastoma by releasing our consensus model, and 3) demonstrate the FL effectiveness at such scale and task-complexity as a paradigm shift for multi-site collaborations, alleviating the need for data-sharing.

Background: Emerging artificial intelligence (AI) technologies have diverse applications in medicine. As AI tools advance towards clinical implementation, skills in how to use and interpret AI in a healthcare setting could become integral for physicians. This study examines undergraduate medical students’ perceptions of AI, educational opportunities about of AI in medicine, and the desired medium for AI curriculum delivery. Methods: A 32 question survey for undergraduate medical students was distributed from May–October 2021 to students to all 17 Canadian medical schools. The survey assessed the currently available learning opportunities about AI, the perceived need for learning opportunities about AI, and barriers to educating about AI in medicine. Interviews were conducted with participants to provide narrative context to survey responses. Likert scale survey questions were scored from 1 (disagree) to 5 (agree). Interview transcripts were analyzed using qualitative thematic analysis. Results: We received 486 responses from 17 of 17 medical schools (roughly 5% of Canadian undergraduate medical students). The mean age of respondents was 25.34, with 45% being in their first year of medical school, 27% in their 2nd year, 15% in their 3rd year, and 10% in their 4th year. Respondents agreed that AI applications in medicine would become common in the future (94% agree) and would improve medicine (84% agree Further, respondents agreed that they would need to use and understand AI during their medical careers (73% agree; 68% agree), and that AI should be formally taught in medical education (67% agree). In contrast, a significant number of participants indicated that they did not have any formal educational opportunities about AI (85% disagree) and that AI-related learning opportunities were inadequate (74% disagree). Interviews with 18 students were conducted. Emerging themes from the interviews were a lack of formal education opportunities and non-AI content taking priority in the curriculum. Conclusion: A lack of educational opportunities about AI in medicine were identified across Canada in the participating students. As AI tools are currently progressing towards clinical implementation and there is currently a lack of educational opportunities about AI in medicine, AI should be considered for inclusion in formal medical curriculum.

Cancer is associated with significant morbimortality globally. Advances in screening, diagnosis, management and survivorship were substantial in the last decades, however, challenges in providing p... Plain language summary Cancer is associated with significant morbimortality globally. Although significant advances occurred in the last decades, challenges in providing personalized care remain. A...

At our institution, we have historically treated brain metastasis (BM) ≤2 cm in eloquent brain with a radiosurgery (SRS) lower prescription dose (PD) to reduce the risk of radionecrosis (RN). We sought to evaluate the impact of this practice on outcomes. We analyzed a prospective registry of BM patients treated with SRS between 2008 and 2017. Incidences of local failure (LF) and RN were determined and Cox regression was performed for univariate and multivariate analyses (MVAs). We evaluated 1533 BM ≤2 cm. Median radiographic follow-up post SRS was 12.7 months (1.4–100). Overall, the 2-year incidence of LF was lower for BM treated with PD ≥21 Gy (9.3%) compared with PD ≤15 Gy (19.5%) (sub–hazard ratio, 2.3; 95% CI: 1.4–3.7; P = 0.0006). The 2-year incidence of RN was not significantly higher for the group treated with PD ≥21 Gy (9.5%) compared with the PD ≤15 Gy group (7.5%) (P = 0.16). MVA demonstrated that PD (≤15 Gy) and tumor size (>1 cm) were significantly correlated (P < 0.05) with higher rates of LF and RN, respectively. For tumors ≤1 cm, when comparing PD ≤15 Gy with ≥21 Gy, the risks of LF and RN are equivalent. However, for lesions >1 cm, PD ≥21 Gy is associated with a lower incidence of LF without significantly increasing the risk of RN. Our results indicate that rates of LF or RN following SRS for BM are strongly correlated with size and PD. Based on our results, we now, depending upon the clinical context, consider increasing PD to 21 Gy for BM in eloquent brain, excluding the brainstem.