Large language models (LLMs) are increasingly transforming medical applications. However, proprietary models such as GPT-4o face significant barriers to clinical adoption because they cannot be deployed on site within healthcare institutions, making them noncompliant with stringent privacy regulations. Recent advancements in open-source LLMs such as DeepSeek models offer a promising alternative because they allow efficient fine-tuning on local data in hospitals with advanced information technology infrastructure. Here, to demonstrate the clinical utility of DeepSeek-V3 and DeepSeek-R1, we benchmarked their performance on clinical decision support tasks against proprietary LLMs, including GPT-4o and Gemini-2.0 Flash Thinking Experimental. Using 125 patient cases with sufficient statistical power, covering a broad range of frequent and rare diseases, we found that DeepSeek models perform equally well and in some cases better than proprietary LLMs. Our study demonstrates that open-source LLMs can provide a scalable pathway for secure model training enabling real-world medical applications in accordance with data privacy and healthcare regulations.

T-cell lymphoblastic lymphoma (T-LBL) and T-cell acute lymphoblastic leukemia (T-ALL) originate from thymic T-cell precursors, with ongoing debate on whether they are variants of the same disease or distinct entities. For 211 patients, including pediatric and adult T-ALL and T-LBL cases, targeted next-generation sequencing and SNP-arrays were performed, and single-nucleotide variants, indels and copy-number variants (CNVs) were analyzed. We aimed to assess genetic differences between T-ALL and T-LBL across age. Generally, mutational landscape analysis identified mutated PHF6 being associated with higher, NOTCH1 with lower age at diagnosis for both T-LBL and T-ALL. Association of CNVs with higher age was evident for T-ALL, but not T-LBL. Analysis of clonal evolution revealed that CNVs – especially deletions and LOH in chromosome 9 (LOH_in_9p) – were observed as first mutational event in both pediatric T-ALL and T-LBL. The sequence of genetic events, starting with LOH_in_9p followed by mutations in NOTCH1, was significantly more frequent in pediatric T-ALL and T-LBL. Detailed evaluation of the patients’ individual clonal evolution indicated that the proportion of malignant cells without NOTCHMT determines the risk of relapse (hazard ratio 1.032, p = 4.65*10−5). In T-ALL, aside from MRD, validated molecular markers for risk-group stratification remain limited. Our data suggest that molecular metrics analogous to those in T-LBL may help refining risk stratification in T-ALL as well.

Background

Large language models (LLMs) are increasingly used by employees at university hospitals for information retrieval or decision support. Self-hosted on-premise systems provide a secure environment and conform to data privacy and security regulations for handling sensitive personal data. Automation of standard procedures using an LLM application can substantially reduce time-consuming administrative tasks and facilitate the analysis of large datasets.

Objective

The objective of our study was to gather feedback from registered artificial intelligence (AI) users on the usability and common use cases of the on-premise LLM infrastructure we established at the University Medicine Magdeburg to optimize the models to the needs of our facility.

Methods:We developed an online questionnaire to which registered AI users were given access and were informed via email.

Results

Of 322 registered AI users, 98 (30.4%) participated in the user survey. After filtering incomplete responses, results from 91 (28.3%) participants remained for further analysis. Speed and quality received overall high approval rates. Most of the users (n=57, 62.6%) used the platform at least once per week, and 44% (n=40) of the users reported saving at least 30 minutes of work per week by using our AI platform. A diverse set of use cases was observed, varying by profession; for example, health care and research professionals used the AI platform more frequently for creation and analysis tasks than administrative staff.

Conclusions

Our data indicate that the implementation of a self-hosted on-premise LLM was associated with positive perceptions among a diverse group of professionals working at a university hospital, saving time and meeting their individual needs.