Stakeholders Want Machine Learning to Support Decisions, Not Replace Them
Machine learning tools for predicting diabetic foot ulcers and amputations will not be adopted on technical merit alone. In this interview study, patients, healthcare workers, and system-integration experts described a narrow set of conditions under which predictive models would feel useful, credible, and fit for routine diabetes care.
The design was intentionally scenario-based. Five diabetes patients, two healthcare workers, and four experts in system integration took part in semi-structured interviews, with three scenario components used to help them articulate what they wanted from predictive tools. Their answers clustered around practical clinical use rather than abstract enthusiasm for artificial intelligence. The most persuasive functions were those that supported judgment, reduced friction in workflows, or delivered an alert before deterioration became visible at the bedside.
Four Use Cases Emerged From the Interviews
Participants identified four concrete scenarios for model use. The first was supporting healthcare workers’ decision-making. In this framing, the model acts as a second layer of clinical intelligence, surfacing risk estimates that can sharpen attention during diabetic foot assessments and follow-up visits.
A second scenario focused on patient empowerment. Here the model is not just an internal tool for clinicians. It becomes a communication aid that can help patients understand why their risk is rising and why adherence, foot care, or timely review matters.
A third use case centered on appointment planning based on risk. In a disease where delayed recognition can lead to tissue loss, model-driven stratification could help clinics prioritize high-risk patients and match resources to clinical need.
The fourth scenario involved early warnings about acute risk. This is where predictive modeling becomes most operationally urgent: an alert that a patient is moving toward a foot ulcer or amputation risk state could trigger faster review, sharper surveillance, or escalation before the window for prevention narrows.
Early warnings were valued most when they could change action, not just generate data.
Explainability and Workflow Fit Mattered as Much as Accuracy
Across the interviews, thirteen values were identified, and many pointed to the same underlying expectation: the model has to fit real care processes. Participants wanted explainability, interoperability, privacy, and regulatory compliance, but they also wanted something less technical and harder to engineer—confidence that the model would feel unobtrusive and clinically relevant rather than disruptive.
The need to translate patient data from electronic health record systems into usable clinical insight came through clearly. Raw data streams are abundant in modern diabetes care, but abundant data do not automatically produce useful decisions. A model that converts those records into risk-based stratification has greater chance of being accepted because it connects prediction to an operational next step.
Integration into existing systems was another recurring requirement. A model that lives outside the clinic’s digital infrastructure risks becoming a parallel workflow, and parallel workflows routinely fail under real-world time pressure. Interoperability was therefore not a technical afterthought but a condition for use.
Patients and Clinicians Wanted a Hybrid Model of Care
One of the strongest themes was the preference for a hybrid approach. Participants did not describe a future in which the algorithm replaces the clinician or the patient. They described a layered system in which machine learning augments human judgment, supports collaboration between healthcare workers at different levels, and improves the timing of intervention.
That preference fits the clinical reality of diabetic foot disease. Risk is not static, and the pathway from neuropathy or skin breakdown to ulceration or amputation can accelerate quickly. A model that flags change early may help clinicians intervene before symptoms become severe, but only if it is placed within a broader care pathway that includes review, communication, and follow-up.
The emphasis on collaboration between healthcare workers at different levels is especially relevant in diabetes care, where nurses, physicians, podiatrists, and other team members often share responsibility. A predictive signal that can be interpreted across roles may have more value than a model designed for a single user group.
Privacy, Compliance, and Continuous Development Were Non-Negotiable
Several values extended beyond bedside utility into governance. Privacy and regulatory compliance were not optional extras. They were part of what made the tool acceptable at all. This is particularly important for models that rely on routine electronic health record data, since the promise of early prediction depends on broad data access, while trust depends on careful limits around that access.
Continuous development also emerged as a necessary attribute. Predictive models are not static interventions; their performance can drift as patient populations, documentation practices, and clinical workflows change. Participants appeared to recognize that a model used for diabetic foot risk will need ongoing refinement if it is to remain clinically relevant.
Unobtrusiveness mattered for the same reason. If the system interrupts too often, overwhelms users, or produces alerts without clear purpose, it can become background noise. In a busy diabetes clinic, noise is fatal to adoption.
Why These Preferences Matter for Diabetic Foot Prediction
Predicting diabetic foot ulcers and amputations is not only a data science challenge. It is a workflow, communication, and trust problem. The study shows that stakeholders were willing to consider machine learning when it supported concrete clinical tasks: decision-making, patient education, appointment prioritization, and early warning.
The most important implication is not that a predictive model must be accurate in isolation. It must be explainable enough to trust, integrated enough to use, and specific enough to change care. A black-box score that never reaches the right clinician at the right time, or cannot be translated into action, offers little benefit no matter how sophisticated the underlying algorithm may be.
By linking stakeholder values to four use scenarios, the interviews provide a practical framework for development. Model builders can use those preferences to shape interface design, alert logic, data integration, and implementation strategy. In diabetes care, where foot complications can progress silently before they become visible, the margin for poor design is small.
The study was based on a small interview sample, and the perspectives came from five patients, two healthcare workers, and four integration experts. Even so, the findings point toward a clear development principle: predictive analytics for diabetic foot complications will be adopted more readily when they help people act earlier, communicate better, and work within existing clinical systems rather than around them.