People with Intellectual Disabilities (ID) are over-represented in forensic and criminal justice settings, but remain under-served by existing static and actuarial measures of risk that are geared toward more general offender populations. These tools do not capture the dynamic, contextual and cognitive nature of the former population, while newer machine learning techniques are criticized for functioning as a black box. We develop a multimodal risk assessment framework based on Explainable Artificial Intelligence, which takes into account heterogeneous data streams - structured clinical and criminological data, behavioral observation time series, neuropsychological and psychometric assessment, and narrative text - to simultaneously predict offending behavior and treatment success in offenders with ID. Our framework consists of modality-specific encoders, adaptive fusion layers that are capable of taking advantage of complementary information, and a temporal self-attention mechanism to model longitudinal trajectories. We bake interpretability into our framework from the very beginning as opposed to treating it as an afterthought and leverage global additive attributions, local Shapley values, and counterfactuals to generate an intelligible explanation of each prediction. This paper also takes into account the issues of class imbalance and fairness of algorithms because recidivism or re-offending cases are less frequent compared to non-recidivism or non-re-offending cases, and there is also an uneven distribution of impact due to the vulnerable nature of the group involved. The paper achieves this by providing the scientific basis for multimodal fusion through the integration of previous empirical research findings, providing a replicable process of evaluation, and identifying the governance requirements needed for deployment of the model.