Prediction of Diabetes Among Homeless Adults Using Artificial Intelligence: Suggested Recommendations
Summary: Published March 22, 2026 in Healthcare (MDPI), this case-control study from Cairo University Faculty of Nursing, Beni-Suef University, and Saudi Electronic University applies machine learning-based diabetes prediction to a medically underserved population — homeless adults — using a hybrid stacking ensemble model. Led by Khadraa Mohamed Mousa and Manal Mohamed Elsawy (Community Health Nursing, Cairo University), the study enrolled 150 homeless adults at the Ma’ana Rescue Human Foundation (Giza, Egypt), including 99 confirmed diabetic cases (FBS ≥ 126 mg/dL or prior diagnosis) and 51 non-diabetic controls. Structured interviews collected demographic data, medical history, 15-item lifestyle questionnaire, and 7-item diabetes knowledge assessment; physiological measurements included weight, height, BMI, waist circumference, triceps skinfold thickness (TSF), fasting blood sugar, and blood pressure. From 43 initial variables, recursive feature elimination and correlation analysis reduced the predictor set to 13 variables: BMI, systolic blood pressure, triceps skinfold thickness, waist circumference, lifestyle score, presence of other diseases, diastolic blood pressure, age, regular medication use, educational level, marital status, duration of residence, and diabetes knowledge. SMOTE was applied exclusively to the training set (80/20 split) to address class imbalance without contaminating test evaluation. Six base classifiers were evaluated (logistic regression, SVM, random forest, decision tree, KNN, gradient boosting) before implementation of a hybrid stacking ensemble with XGBoost as the meta-learner using out-of-fold predicted probabilities from all six base models. The stacking ensemble achieved 95.45% accuracy, 100% precision, 93.75% recall, F1-score 0.968, and AUC-ROC 0.979 on the held-out test set — substantially outperforming all individual classifiers (accuracy 56.7–70%, F1 0.686–0.781). Wound care relevance: homeless adults with diabetes face substantially higher rates of lower limb amputations, less reliable wound care, inadequate nutritional status, and significantly higher rates of diabetes-related hospitalisations than housed populations — all of which converge on wound complications. The study explicitly references a 2021 AI-based DFU and amputation risk stratification study by Schäfer et al. as the broader clinical context. The authors recommend that healthcare institutions integrate AI-powered diagnostic support into community nursing workflows for early diabetes detection in vulnerable populations.
Key Highlights:
- Stacking ensemble performance: hybrid XGBoost meta-learner achieved 95.45% accuracy, 100% precision, AUC 0.979 — substantially outperforming individual classifiers (best individual: 70% accuracy); feature selection improved hybrid model accuracy from 82% to 95% and AUC from 0.87 to 0.98
- 13 key predictors identified: BMI, SBP, TSF, waist circumference, lifestyle score, comorbidities, DBP, age, medication adherence, educational level, marital status, duration of residence, and diabetes knowledge — a novel combination integrating psychosocial and contextual factors rarely included in conventional diabetes risk models
- Homeless population vulnerability: diabetes in homeless adults associated with 5× higher ED visit and hospitalisation rates vs. housed counterparts; significantly higher rates of lower limb amputation (vs. 0.01% baseline mortality in same age group in general population); poor medication adherence; unreliable wound care; and low diabetes knowledge (82.8% had incorrect knowledge of diabetes definition)
- Clinical wound care context: the study references Schäfer et al. (2021, Front Med) on ML-based DFU and amputation risk stratification as its broader framework — positioning early AI-assisted diabetes detection in homeless populations as an upstream prevention strategy for the DFU and amputation pipeline
- Limitations: single-centre, n=150, purposive sample; case-control design reflects institutional prevalence rather than community prevalence; small test set (n=30) may inflate performance estimates; external validation in larger multi-centre samples is required before clinical deployment
- Nursing recommendation: community and gerontological health nurses are positioned to implement AI-assisted screening alongside fingertip glucose testing in shelter and community settings — providing instant results and enabling same-encounter lifestyle counselling for high-risk homeless adults
Keywords: diabetes prediction machine learning, homeless population diabetes wound risk, AI diabetic foot amputation risk, community nursing diabetes screening, ensemble machine learning healthcare, health equity diabetes vulnerable population
Khadraa Mohamed Mousa, Farid Ali Mousa, Naglaa Mahmoud Abdelhamid, Mona Sayed Atress, Manal Mohamed Elsawy