A mature business lady is doing a phone call, holding her smartphone...

Article • Promising technology for early detection

Phonebots listen out for Alzheimer’s

Dementia is one of Europe’s most pressing health and societal challenges, and early detection of Alzheimer’s disease (AD) is becoming more important as new treatments emerge. At the joint annual congress of the German Society for Geriatrics (DGG) and the German Society for Gerontology and Geriatrics (DGGG) in Frankfurt am Main in September, Dr Stefanie Köhler presented “Mili”, an automated speech-based phonebot designed to support the early detection of AD. Her findings show that the technology is usable, but that meaningful consent, privacy, fairness and human support must remain central to its use.

Article: Cornelia Wels-Maug

Dementia is one of Europe’s most significant health and societal challenges, and the number of people affected is expected to rise considerably in the coming decades. Alzheimer Europe estimates that 12.1 million people lived with dementia in Europe (EU27 and non-EU countries combined) in 2025, a figure projected to rise to 19.9 million by 2050.¹ Worldwide, more than 55 million people live with dementia, a number believed to rise to 139 million by 2050.² Alzheimer’s disease accounts for an estimated 60–70% of dementia cases.³

The treatment landscape changed in 2025, when the European Commission approved two drugs for treating early symptomatic AD: lecanemab (Leqembi) and donanemab (Kisunla). Both slow cognitive decline in people with mild cognitive impairment (MCI) or mild dementia. Because these treatments are intended for early symptomatic disease, identifying people who may be developing AD before they show severe symptoms has become important. This creates a need for accessible, scalable screening tools that can flag individuals who should be referred for further clinical assessment.

Traditionally, cognitive decline is diagnosed through in-person neuropsychological tests administered by a clinician. This time-consuming procedure does not lend itself to large-scale screening, which is why speech as a biomarker looks promising: it is relatively easy to collect, non-intrusive and can be captured remotely. Advances in artificial intelligence (AI), machine learning and natural language processing make it possible to analyse speech data faster and potentially at scale.⁴ However, many studies on speech as a biomarker of AD have incomplete datasets, which makes it difficult to train algorithms reliably. Researchers are therefore working to build large, standardised and scalable databases for speech biomarker research.

Positive user feedback for Mili

Köhler and colleagues used a user-centred design approach to evaluate and improve the usability of phonebot Mili, with the aim of meeting the needs of people with cognitive impairment.⁵ The study, published in Alzheimer's & Dementia, included 234 participants, comprising healthy older adults and people with cognitive impairment. Using a female voice, Mili automatically called participants every three months over a 15-month period. Each call lasted about 15 minutes and included three cognitive tasks: word-list recall, semantic verbal fluency and storytelling.

Usability was assessed with the System Usability Scale (SUS), which yielded a score of 70.4 out of 100. Köhler also conducted semi-structured interviews to explore participants’ experiences, their confidence in the technology and their suggestions for improvement. Overall, participants found the phonebot easy to use, although some had to overcome an initial learning curve, and many expressed a desire for access to technical support. Participants tended to anthropomorphise the technology, which blurred the boundary between human and automated interaction.

Some interviewees said they would use Mili regularly if they received feedback on their cognitive performance, although they were more interested in learning how to prevent or reduce their risk of AD. Köhler also noted that participants would appreciate appropriate guidance when results suggest possible cognitive decline. Participants generally felt that Mili should complement, rather than replace, a professional clinical evaluation.

Ethical considerations

Beyond these usability findings, the use of automated phonebots for speech-based cognitive screening raises important ethical questions, including obtaining meaningful consent, protecting privacy, minimising psychological distress, ensuring fairness and preserving access to human care.

Individuals with MCI or AD can experience fluctuating decision-making capacity, which may affect their ability to understand the purpose and implications of taking part in data collection. Participants may also mistake an automated phonebot for a healthcare professional, or assume that the interaction constitutes treatment rather than research or screening. One-time consent models may therefore be insufficient. Researchers should consider ongoing consent procedures that periodically confirm participants’ understanding of the process and their willingness to continue.

Speech recordings can reveal personal identity and sensitive information. During unstructured and semi-structured conversations, participants may unintentionally disclose medical, financial or family issues beyond the intended scope of the conversation. Repeated calls may raise concerns about surveillance, and shared devices, background conversations and the involvement of family members pose additional privacy risks. Researchers must clearly explain what is recorded, how data are stored and used, who can access them, and how confidentiality is protected.

Automated cognitive assessments can trigger performance anxiety, confusion or feelings of failure, particularly when participants struggle to find words or misunderstand instructions. Speech recognition errors can also cause frustration. If a phonebot identifies signs of cognitive decline, careful consideration must be given to how these findings are communicated. Human support is advisable when results may be concerning, so that context and psychological support can be provided to minimise anxiety or distress.

Algorithmic fairness, transparency and bias

Speech-analysis systems can perform unevenly across populations if their training data lack diversity. Regional accents, non-native speech, hearing loss and communication disorders could lead to inaccurate classifications and inequitable screening outcomes. Algorithms should therefore be validated across diverse groups, and results should not be treated as definitive diagnoses.

Deep-learning models can also be difficult to interpret, which makes it challenging to explain why a particular result was produced. Researchers should communicate the system’s capabilities, limitations and uncertainties in simple language. There is also a risk that automated screening could replace valuable clinical contact and personal support, leaving older adults feeling isolated or unsupported.

Ultimately, phonebots offer a promising way to expand access to speech-based cognitive screening. However, their use must be guided by meaningful consent, robust data protection, algorithmic fairness and appropriate safeguards for participants’ psychological well-being. Crucially, automated screening should complement rather than replace professional clinical assessment and human support. Only with these measures in place can phonebots be developed and used responsibly.

Profile:

Dr Stefanie Köhler is a researcher at the German Center for Neurodegenerative Diseases (DZNE) Rostock/Greifswald, Germany, and is affiliated with the Department of Psychosomatic Medicine at Rostock University Medical Center. Her research focuses on speech and digital biomarkers for the early detection of Alzheimer’s disease, including the automated phonebot Mili.

References:

02.11.2026