<![CDATA[Smartphone cameras and lightweight deep learning (DL) enable scalable movement analysis for post–lower-limb-injury rehabilitation; however, the scientific paradigms that guide model design and evaluation remain under-discussed. This study compares Positivist and Post-Positivist paradigms and formulates a techno-humanistic methodological framework for developing lightweight DL rehabilitation assessment systems. The core methodology integrates (i) bibliometric analysis of 35 publications (2020–2025) on lightweight DL, computer vision, explainable AI, and clinical rehabilitation, and (ii) philosophical reflective comparative analysis across ontology, epistemology, and axiology. Using VOSviewer network, overlay, and density visualizations, we map keyword co-occurrence structures and interpret their methodological orientation through the two paradigms. The results show a dominant Positivist-oriented technical cluster emphasizing measurable performance and reproducible optimization (e.g., classification, accuracy, real-time and edge deployment). In parallel, emerging Post-Positivist themes highlight external validity, interpretability (XAI), fairness, demographic representativeness, and quality-of-life outcomes, indicating a shift toward context-aware and ethically grounded evaluation. By combining bibliometric mapping with philosophical synthesis, this framework provides a reproducible methodological lens for informatics/computer science researchers to design lightweight AI that is not only efficient, but also explainable, clinically relevant, and ethically accountable. ]]>
