Existing machine learning (ML) models for dementia detection typically use fixed prediction windows and often lack clinically informed variable selection and responsible AI principles, thereby limiting performance, generalisability and clinical trust. We aimed to develop a single clinically informed model capable of both diagnosing dementia and predicting its onset over flexible time horizons up to 10 years. We trained our ML model on the NACC Uniform Data Set using 20 inputs compatible with UK National Healthcare Service secondary-care workflows, including a HORIZON feature encoding months from baseline. We also integrated explainability and fairness analyses. Same-visit classification achieved sensitivity of 0.951 (95% confidence interval 0.947–0.954), specificity of 0.810 (0.805–0.814), a geometric mean (G-mean) of 0.877 (0.874–0.880) and an area under the receiver operating characteristic curve (AUC) of 0.959 (0.957–0.960). Across future horizons from 24 to 120 months, sensitivity/specificity ranged between 0.808/0.856 at 24 months and 0.833/0.706 at 120 months, while AUCs were 0.906 and 0.851. Compared with a Cox proportional hazards model, our flexible-horizon ML model achieved a higher G-mean at four of the five evaluated horizons and higher specificity from 48 months onwards, whereas Cox achieved greater sensitivity from 48 months onwards. Shapley additive explanations highlighted functional independence, prediction horizon and cognitive measures as influential. Exploratory subgroup analyses showed broadly similar balanced discrimination across sex, race and age groups, although sensitivity–specificity trade-offs varied. An interactive web application was also developed for demonstration and clinical feedback. These findings support the feasibility of flexible-horizon, clinically-informed and responsible ML-based decision support for dementia early detection; however, external validation, recalibration and prospective clinical evaluation are required before deployment.