DialysisBot: AI Chatbot for Hemodialysis Self-Management

Background and Rationale

Patients with end-stage renal disease undergoing chronic hemodialysis must manage a demanding interdialytic regimen, including fluid restriction, dietary control of phosphorus, potassium, and sodium intake, and vascular access care (arteriovenous fistula or central venous catheter). The cognitive and behavioral burden of this regimen, combined with limited continuous support outside dialysis sessions, contributes to often suboptimal adherence. Elevated interdialytic weight gain, hyperphosphatemia, and hyperkalemia are well-documented consequences of non-adherence and are associated with increased morbidity.

mHealth technologies and AI-based conversational agents (chatbots) have shown growing potential to support patients with chronic conditions by improving access to personalized health information, increasing patient engagement, and promoting appropriate self-management behaviors, particularly in populations with limited access to traditional educational resources. Systematic reviews have reported positive effects of conversational agents on clinical and behavioral outcomes in chronic disease, and a recent scoping review mapping mHealth use among dialysis patients found that self-management applications were associated with improvements in interdialytic weight gain, phosphatemia, potassium levels, and adherence to dietary prescriptions.

Despite this promising evidence base, data specific to AI-based chatbots in the hemodialysis population remain limited. To date, neither the international literature nor Italian clinical practice offers a validated conversational agent specifically designed for this population; the gap concerns both the lack of purpose-built solutions and the absence of systematic feasibility, acceptability, and impact evaluations in real-world clinical settings. This study addresses this gap by implementing and evaluating the clinical feasibility, safety, and impact of DialysisBot, hypothesizing that such an intervention may contribute to patient empowerment, reduce interdialytic complications, and improve overall quality of care.

Study Objectives

Primary objective: to evaluate the feasibility, acceptability, impact, and satisfaction associated with the use of an AI-based conversational agent (DialysisBot) to support self-management among patients undergoing hemodialysis.

Secondary objectives:

• To assess support for dietary self-management, with particular attention to phosphorus, potassium, and sodium restriction.
• To assess support for fluid intake control during the interdialytic period.
• To assess whether the guidance provided by the system on vascular access management (arteriovenous fistula and central venous catheter) is implemented by patients and consistent with current standards of care.
• To assess the intervention's impact on treatment adherence.
• To explore barriers, facilitators, and overall user experience reported by hemodialysis patients.

Study Design

This is a prospective single-arm interventional pilot study organized in two methodological phases. Phase 1 is a quantitative longitudinal phase covering three assessment time points (T0-T2), using standardized questionnaires and routine clinical parameters. Phase 2 is a qualitative phase conducted after T2, based on semi-structured interviews analyzed through reflexive thematic analysis following the Braun and Clarke (2006) approach.

Assessment time points:

• T0 (Baseline): participant enrollment, informed consent, administration of baseline assessment instruments, and application training.
• T1 (1 month after initiation): first interim assessment, including administration of the usability scale and collection of routine clinical parameters.
• T2 (3 months after initiation): final assessment, including all outcome instruments, clinical data, and, for the qualitative subgroup, semi-structured interviews.

Intervention Description: DialysisBot

DialysisBot is a multiplatform, AI-based chatbot application designed to support hemodialysis patients in the day-to-day management of their condition. The system uses a hybrid architecture combining three components: (1) a large generative language model responsible for semantically rephrasing patient questions and managing natural dialogue; (2) a fine-tuned BERT classifier that identifies the domain of each request (diet, fluid intake, vascular access management, or organizational aspects of dialysis treatment); and (3) a vector-similarity retrieval engine that controls access to validated educational content stored in the study's clinical database.

The system operates exclusively on a clinically validated document base, drawing solely on materials preloaded by the research team, corresponding to guidelines and protocols approved by the Istituto Superiore di Sanità (ISS) and the local hospital institution. The system performs no autonomous web search, and uncontrolled external sources are entirely outside its scope.

The core safety mechanism compares each user query against indexed reference documents using cosine similarity. When the similarity score falls below a predefined threshold - empirically optimized during testing with real users - the system withholds generation of a clinical response and instead returns a default message informing the patient that it cannot answer, directing them to healthcare staff. This mechanism is the main technical safeguard against responses disconnected from validated sources, i.e., against hallucinations.

To orient users and reduce out-of-scope queries, the interface offers a set of guiding questions illustrating the types of requests the system can handle (e.g., "How much potassium is in an apple?", "How can I protect my fistula?", "What are the early signs of a CVC infection?"). Conceived as a conversational educational environment, the interface returns personalized, contextual information consistent with available clinical recommendations. The system does not formulate diagnoses or suggest treatment decisions; responses are limited to retrieval within the uploaded informational and educational content, accompanied by safety messages and, where appropriate, a prompt to contact healthcare staff.

Participant Training

Before the start of the study, all participants receive an individual training session led by dedicated healthcare staff, including: (a) a guided walkthrough of the application interface; (b) an integrated digital user manual accessible within the app at any time; and (c) ongoing technical support from the research team throughout the study. Training preferably takes place during a routine dialysis session to minimize additional burden on the patient, or is delivered via chat for patients recruited online.

Outcomes and Assessment Instruments

Instrument selection was informed by recent literature on mHealth technologies in dialysis patients and by patient-reported outcome measures (PROMs) recommended in nephrology.

Primary outcomes:

• Feasibility, assessed through: recruitment rate (proportion of eligible patients who consent to participate); retention rate (proportion of participants completing the study at T2); questionnaire completion rate (proportion of instruments adequately completed from T1 to T2); and application engagement (mean number of weekly interactions per patient, derived from anonymized system logs).
• Usability and user experience, assessed using the Italian version of the Chatbot Usability Scale (BUS-11), an 11-item, 5-point Likert scale validated for AI conversational systems (total score range 11-55; higher scores indicate greater perceived usability).
• Treatment adherence, measured through clinical outcome indicators collected at T0, T1, and T2: interdialytic weight gain (IDGW from pre- and post-dialysis weight); serum phosphate (mmol/L) and serum potassium (mmol/L) from routine blood draws; and dialysis session attendance rate (%) from dialysis center attendance records.

Secondary outcomes:

• Patient-reported outcomes, assessed using the Italian version of the Kidney Disease Quality of Life instrument (KDQOL-ita 1.3, 36 items), a validated, Italian-translated instrument recommended as a reference PROM in nephrology (subscale scores range 0-100; higher scores indicate better health-related quality of life).

Qualitative Data Collection

In a purposively selected subgroup of participants (n = 10-15, chosen to maximize variability in age, gender, and digital literacy), semi-structured interviews will be conducted at the end of the study (T2), as part of the qualitative Phase 2. This phase aims to explore participants' lived experience of using DialysisBot. Interviews will be audio-recorded, transcribed verbatim, and analyzed through reflexive thematic analysis according to Braun and Clarke (2006), with independent coding by two researchers and calculation of inter-rater agreement (Cohen's kappa). Thematic areas explored include: overall experience with the application; perceived impact on self-management behaviors; barriers and facilitators to use; level of trust in the information provided by the system; and areas for improvement.

Study Timeline

The project is organized into five operational phases: Phase 1 (1-2 months), application and scientific content design; Phase 2 (1-2 months), technical development and internal alpha/beta testing; Phase 3 (baseline/T0), enrollment, consent, baseline data collection, and participant training; Phase 4, comprising the T1 interim assessment at 1 month (usability and adherence questionnaires) and the T2 final assessment at 3 months (all outcome instruments, clinical data, and interviews); and Phase 5 (2 months), data analysis and reporting of results.

Data Analysis

Quantitative analysis: descriptive statistics will be calculated (means and standard deviations for continuous variables, or percentage frequencies for sociodemographic and clinical sample characteristics). Pre-post comparisons (T0 vs. T2) will be conducted using the paired-samples t-test for normally distributed variables (assessed with the Shapiro-Wilk test), or the Wilcoxon signed-rank test for non-normally distributed variables. The significance threshold will be set at p < 0.05. Analyses will be performed using Jamovi software.

Qualitative analysis: thematic analysis will follow the Braun and Clarke (2006) framework, comprising familiarization with the data, generation of initial codes, searching for themes, reviewing themes, defining and naming themes, and producing the final report. Coding will be performed independently by two researchers, with inter-rater agreement calculated using Cohen's kappa. Qualitative findings from Phase 2 will be presented in narrative and thematic form.

AI System Safety and Governance

The clinical and research team is multidisciplinary and includes: the Principal Investigator; a PhD candidate and a hemodialysis nurse (member of the Italian Society of Nephrology Nursing, SIAN), responsible for study coordination and data collection; an AI supervisor, an Associate Professor with expertise in applied artificial intelligence, responsible for oversight of the system architecture and technical validation; and a doctoral-level nursing supervisor, responsible for ensuring that the information provided by the system is consistent with current clinical guidelines and protocols.

DialysisBot was designed in accordance with the ethical principle of non-maleficence and with the European regulatory framework for AI systems (EU AI Act, 2024). The system is classifiable as a limited-risk AI application (an informational chatbot), subject to user transparency obligations.

Key safety and governance measures include: informational content generated exclusively from validated sources approved by the clinical team; no autonomous web search or use of uncontrolled external sources; for questions the system cannot handle, explicit communication of this limitation and referral to healthcare staff; anonymized logging and periodic team review of unresolved interactions, to progressively improve content coverage; clear and visible safety warnings reminding users of the tool's informational and educational nature, which does not replace the clinical judgment of healthcare professionals; and periodic clinical review of content by the research team.

Data protection and security: user conversations are stored in pseudonymized form, with each participant assigned a unique identification code disconnected from personally identifiable information. Text data is encrypted using the AES-256 (Advanced Encryption Standard, 256-bit) algorithm before being written to the database. The storage server is physically located within the European Union, in compliance with GDPR provisions on international data transfers (Articles 44-49) and with the requirements of the Italian Data Protection Authority.

Ethical Considerations

The study will be conducted in full compliance with the principles of the Declaration of Helsinki (2013 revision) and applicable personal data protection regulations (EU Regulation 2016/679, GDPR). All participants will provide informed consent prior to enrollment. Participation is voluntary, and all collected data will be anonymized or pseudonymized and used exclusively for scientific research purposes. The study protocol has been approved by Ethics Committee No. 2992/CEL, Oncologico Bari.