TWTobias Welt

Tobias Welt

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Tobias Welt has a total of 151 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Top Cited Papers
    Open Access

    Deep learning using electroencephalogram (EEG) data for diagnosing and predicting SSRI response in major depressive disorder

    by Sebastian Olbrich, Natalia Jaworska, Sara de la Salle, Verner Knott, Pierre Blier, Martin Brunovsky, Tobias Welt, Mateo de Bardeci, Cheng Teng-Ip

    Background: Major Depression (MDD) is a potentially life-threatening condition that ranks among the diseases with the highest global burden. Despite its prevalence, current diagnostic methods remain largely subjective, and first-line treatments exhibit high rates of non-responders. Methods: This study investigates the application of deep learning (DL) algorithms to electroencephalogram (EEG) data for the MDD-diagnosis and prediction of treatment outcomes following the administration of selective serotonin reuptake inhibitors (SSRIs), using six large, independent datasets with a total of n = 146 for healthy subjects and n = 203 for patients. DL models were trained on one portion of the datasets and tested on unseen data from different subjects. To interpret the classification features, Gradient-weighted Class Activation Mapping (Grad-CAM) was applied. Results: The models achieve an average accuracy of 67.5% (best fold 70%) in distinguishing MDD patients from healthy controls and mean 79% accuracy (best fold 85%) in predicting SSRI responders. Key EEG markers for both classification tasks revealed by Grad-CAM include alpha activity in the frontal and parietal regions. Simulation of a clinical decision scenario for SSRI treatment selection indicates a number needed to treat (NNT) of five when using a model with 80% predictive accuracy, corresponding to an increase in treatment response from a 50% baseline to 70% with model-guided selection.Conclusion: These findings underscore the clinical potential of EEG-based DL models for stratified treatment in MDD, facilitating accurate therapy choices and reducing ineffective treatments. The results of the integration of objective neurophysiological markers into clinical psychiatry are indicating the potential for more personalized treatment allocation.


  • Open Access

    Can Conversational AI Reliably Assess Depressive Symptoms? A Feasibility, Acceptability and Bias-Aware Validation Study of MADRS-Chat

    by Samantha Weber, Barbora Provaznikova, Nikolas Psathakis, Andrea Casanova, Lucian Amthor, Lara Kirchhofer, José Antonio García Macías, Tobias Welt, Golo Kronenberg, Tobias Kowatsch, et al.

    Background: Structured clinical interviews such as the Montgomery-Åsberg Depression Rating Scale (MADRS) are central to depression assessment but are time-intensive and inconsistently applied in routine care. Large language models (LLMs) offer new opportunities to support standardized symptom assessment through conversational interfaces, yet their clinical feasibility, validity, and potential biases remain insufficiently studied.Methods: We evaluated MADRS-Chat, an LLM-assisted text-based chatbot designed to administer the MADRS interview and generate symptom-level scores using a clinician-trained scoring model. In a prospective observational study involving psychiatric in- and outpatients, we evaluated feasibility and acceptability of a conversational chatbot and its agreement with clinician-rated MADRS scores and patient self-rated (MADRS-S) scores.Findings: MADRS-Chat was well accepted by patients and clinicians and chatbot-derived MADRS scores showed meaningful agreement with human ratings. Qualitative feedback indicated that the anonymous nature of the written interaction facilitated more open and detailed symptom disclosure. However, chatbot-administered interviews systematically yielded higher severity scores, likely reflecting interaction-induced shifts in response style. This bias was effectively addressed through lightweight post-hoc calibration.Interpretation: Our findings demonstrate that conversational AI can support structured psychiatric assessment in real-world clinical settings when used within clearly defined clinical boundaries. Moreover, text-based conversational interactions can alter symptom reporting, making explicit validation and calibration essential before deployment.

  • Top Cited Papers
    Open Access

    Suicide- and crisis-risk detection using large language models in mental-health chatbots

    by Samantha Weber, Dustin Klebe, Lukas Wolf, Christopher Aeberli, Stephanie Homan, Nikolas Psathakis, Andrea Casanova, José Garcia Macias, Jack Sykstus, Jie-Ming Li, et al.

    Objective: Large language models (LLMs) are increasingly embedded in mental-health chatbots, yet safe deployment is limited by two unresolved challenges: (1) suicide- and crisis-risk detection lacks a definitive ground truth and is characterized by substantial clinician disagreement, and (2) most evaluations frame risk detection as an offline accuracy task rather than a real-time safety problem. This study aimed to empirically characterize these limitations and to derive design principles for uncertainty-aware, safety-oriented crisis detection in conversational artificial intelligence. Methods and Analysis: We curated a clinician-labeled dataset of 200 real-world conversation segments drawn from a deployed mental-health chatbot. Five clinical experts independently annotated each segment for suicide- and crisis-related risk. Using a single base LLM, we implemented five prompt-defined detection variants with systematically increasing sensitivity thresholds, without task-specific training or fine-tuning. Models were evaluated against clinician consensus labels to quantify false-negative and false-positive trade-offs. Latency analyses assessed feasibility for real-time, per-turn monitoring. Results: As sensitivity increased, the false-negative rate decreased monotonically from 87% to 0%, while false-positive rates rose accordingly. High - and extreme-sensitivity variants achieved near-perfect (98.9%) and perfect (100%) recall, demonstrating that near–zero–miss crisis detection from natural language is technically feasible in real time (mean latency emergency mode in which conservative risk detection operates independently from the conversational model, allowing supportive engagement to be maintained under heightened safety constraints. This layered, uncertainty-aware architecture provides a practical pathway for safer deployment of LLM-based mental-health chatbots without reliance on large training datasets or extensive model optimization. What is already known on this topic?: Mental-health chatbots based on large language models are increasingly used for psychological support, but evaluations and real-world incidents show that general-purpose LLMs are unreliable in recognizing suicide- and crisis-related risk and may respond unsafely in high-stakes situations. Suicide and crisis risk assessment lacks a stable ground truth, with substantial inter-clinician disagreement even among trained experts, indicating that automated risk detection is inherently uncertain and cannot be treated as a conventional supervised classification task. What this study adds: Demonstrates that near-zero-miss suicide- and crisis-risk detection is technically feasible using clinician-validated data and prompt-based sensitivity calibration, without the further need for fine-tuning or large task-specific training datasets. Shows that detection errors are largely driven by irreducible uncertainty rather than model failure, as misclassifications systematically align with areas of clinician disagreement, supporting the view that crisis detection is an online monitoring problem rather than a solvable classification task. Introduces an architectural safety framework for mental-health chatbots in which conservative, independent risk detection enables an operational emergency mode that prioritizes safety while maintaining empathic engagement. How this study might afect research, practice or policy: This work reframes suicide- and crisis-risk detection in conversational AI as a safety-oriented, uncertainty-aware problem rather than an accuracy-driven prediction task, challenging prevailing evaluation practices. By proposing an architectural separation between risk detection and dialogue generation, it provides a practical, scalable framework for deploying mental-health chatbots that support just-in-time safety interventions without relying on extensive clinical training data.


  • Open Access

    Remission of auditory verbal hallucinations following neuronavigated, personalized rTMS guided by EEG, MRI, and auditory evoked potentials: A case report

    by Barbora Provaznikova, Anna Monn, Aygün Ertuğrul, Samantha Weber, Erich Seifritz, Johannes Jungwirth, Golo Kronenberg, Sebastian Olbrich, Tobias Welt

  • A human-derived antibody targets misfolded SOD1 and ameliorates motor symptoms in mouse models of amyotrophic lateral sclerosis

    by Marcel Maier, Tobias Welt, Fabian Wirth, Fabio Montrasio, Daniel Preisig, Jordan McAfoose, Fernando G. Vieira, Luka Kulic, Claudia Späni, Thilo Stehle, et al.

    Mutations in the gene encoding superoxide dismutase 1 (SOD1) lead to misfolding and aggregation of SOD1 and cause familial amyotrophic lateral sclerosis (FALS). However, the implications of wild-type SOD1 misfolding in sporadic forms of ALS (SALS) remain unclear. By screening human memory B cells from a large cohort of healthy elderly subjects, we generated a recombinant human monoclonal antibody (α-miSOD1) that selectively bound to misfolded SOD1, but not to physiological SOD1 dimers. On postmortem spinal cord sections from 121 patients with ALS, α-miSOD1 antibody identified misfolded SOD1 in a majority of cases, regardless of their SOD1 genotype. In contrast, the α-miSOD1 antibody did not bind to its epitope in most of the 41 postmortem spinal cord sections from non-neurological control (NNC) patients. In transgenic mice overexpressing disease-causing human SOD1G37R or SOD1G93A mutations, treatment with the α-miSOD1 antibody delayed the onset of motor symptoms, extended survival by up to 2 months, and reduced aggregation of misfolded SOD1 and motor neuron degeneration. These effects were obtained whether α-miSOD1 antibody treatment was administered by direct brain infusion or peripheral administration. These results support the further development of α-miSOD1 antibody as a candidate treatment for ALS involving misfolding of SOD1.


  • High-speed video gait analysis reveals early and characteristic locomotor phenotypes in mouse models of neurodegenerative movement disorders

    by Daniel F. Preisig, Luka Kulic, Maik Krüger, Fabian Wirth, Jordan McAfoose, Claudia Späni, Pascal Gantenbein, Rebecca Derungs, Roger M. Nitsch, Tobias Welt


  • Open Access

    Reduced β-amyloid pathology in an APP transgenic mouse model of Alzheimer’s disease lacking functional B and T cells

    by Claudia Späni, Tobias Suter, Rebecca Derungs, Maria Teresa Ferretti, Tobias Welt, Fabian Wirth, Christoph Gericke, Roger M. Nitsch, Luka Kulic

    In Alzheimer’s disease, accumulation and pathological aggregation of amyloid β-peptide is accompanied by the induction of complex immune responses, which have been attributed both beneficial and detrimental properties. Such responses implicate various cell types of the innate and adaptive arm of the immunesystem, both inside the central nervous system, and in the periphery. To investigate the role of the adaptive immune system in brain β-amyloidosis, PSAPP transgenic mice, an established mouse model of Alzheimer’s disease, were crossbred with the recombination activating gene-2 knockout (Rag2 ko) mice lacking functional B and T cells. In a second experimental paradigm, aged PSAPP mice were reconstituted with bone marrow cells from either Rag2 ko or wildtype control mice. Analyses from both experimental approaches revealed reduced β-amyloid pathology and decreased brain amyloid β-peptide levels in PSAPP mice lacking functional adaptive immune cells. The decrease in brain β-amyloid pathology was associated with enhanced microgliosis and increased phagocytosis of amyloid β-peptide aggregates. The results of this study demonstrate an impact of the adaptive immunity on cerebral β-amyloid pathology in vivo and suggest an influence on microglia-mediated amyloid β-peptide clearance as a possible underlying mechanism.


  • Acute Effects of Muscarinic M1 Receptor Modulation on AβPP Metabolism and Amyloid-β Levels in vivo : A Microdialysis Study

    by Tobias Welt, Luka Kulic, Sarah E. Hoey, Jordan McAfoose, Claudia Späni, Antonella Santuccione Chadha, Abraham Fisher, Roger M. Nitsch

    Indirect modulation of cholinergic activity by cholinesterase inhibition is currently a widely established symptomatic treatment for Alzheimer’s disease (AD). Selective activation of certain muscarinic receptor subtypes has emerged as an alternative cholinergic-based amyloid-lowering strategy for AD, as selective muscarinic M1 receptor agonists can reduce amyloid-β (Aβ) production by shifting endoproteolytic amyloid-β protein precursor (AβPP) processing toward non-amyloidogenic pathways. In this study, we addressed the hypothesis that acute stimulation of muscarinic M1 receptors can inhibit Aβ production in awake and freely moving AβPP transgenic mice. By combining intracerebral microdialysis with retrodialysis, we determined hippocampal Aβ concentrations during simultaneous pharmacological modulation of brain M1 receptor function. Infusion with a M1 receptor agonist AF102B resulted in a rapid reduction of interstitial fluid (ISF) Aβ levels while treatment with the M1 antagonist dicyclomine increased ISF Aβ levels reaching significance within 120 minutes of treatment. The reduction in Aβ levels was associated with PKC α and ERK activation resulting in increased levels of the α -secretase ADAM17 and a shift in AβPP processing toward the non-amyloidogenic processing pathway. In contrast, treatment with the M1 receptor antagonist dicyclomine caused a decrease in levels of phosphorylated ERK that was independent of PKC α , and led to an elevation of β-secretase levels associated with increased amyloidogenic AβPP processing. The results of this study demonstrate rapid effects of in vivo M1 receptor modulation on the ISF pool of Aβ and suggest that intracerebral microdialysis with retrodialysis is a useful technical approach for monitoring acute treatment effects of muscarinic receptor modulators on AβPP/Aβ metabolism.


  • Open Access

    Amyloid-β Peptide-specific DARPins as a Novel Class of Potential Therapeutics for Alzheimer Disease

    by Michael Hanenberg, Jordan McAfoose, Luka Kulic, Tobias Welt, Fabian Wirth, Petra Parizek, Lisa Strobel, Susann Cattepoel, Claudia Späni, Rebecca Derungs, et al.


  • Open Access

    Early accumulation of intracellular fibrillar oligomers and late congophilic amyloid angiopathy in mice expressing the Osaka intra-Aβ APP mutation

    by L Kulic, J McAfoose, T Welt, C Tackenberg, C Späni, F Wirth, V Finder, U Konietzko, M Giese, A Eckert, et al.

    Pathogenic amyloid-β peptide precursor (APP) mutations clustered around position 693 of APP—position 22 of the Aβ sequence—are commonly associated with congophilic amyloid angiopathy (CAA) and intracerebral hemorrhages. In contrast, the Osaka (E693Δ) intra-Aβ APP mutation shows a recessive pattern of inheritance that leads to AD-like dementia despite low brain amyloid on in vivo positron emission tomography imaging. Here, we investigated the effects of the Osaka APP mutation on Aβ accumulation and deposition in vivo using a newly generated APP transgenic mouse model (E22ΔAβ) expressing the Osaka mutation together with the Swedish (K670N/M671L) double mutation. E22ΔAβ mice exhibited reduced α-processing of APP and early accumulation of intraneuronal fibrillar Aβ oligomers associated with cognitive deficits. In line with our in vitro findings that recombinant E22Δ-mutated Aβ peptides form amyloid fibrils, aged E22ΔAβ mice showed extracellular CAA deposits in leptomeningeal cerebellar and cortical vessels. In vitro results from thioflavin T aggregation assays with recombinant Aβ peptides revealed a yet unknown antiamyloidogenic property of the E693Δ mutation in the heterozygous state and an inhibitory effect of E22Δ Aβ42 on E22Δ Aβ40 fibrillogenesis. Moreover, E22Δ Aβ42 showed a unique aggregation kinetics lacking exponential fibril growth and poor seeding effects on wild-type Aβ aggregation. These results provide a possible explanation for the recessive trait of inheritance of the Osaka APP mutation and the apparent lack of amyloid deposition in E693Δ mutation carriers.