Table 7 Summary of processing complexity and memory analysis in adaptive and dynamic scheduling.

CCIA⁵⁹

Optimized wavelet transform for low-power IIoT sensors, reducing computational cost and adjust data transmission

Significant energy savings (up to 87%), scalable compression, suitable for batteryless IIoT

Limited to specific MCU platforms; may require assembly-level coding expertise

Radar-PIM⁶⁰

Uses processing-in-memory architecture for efficient UWB radar-based respiration detection on IoT processors

Up to 42% energy savings, 3x performance improvement vs. multicore processors, optimized for respiration detection

Hardware complexity and integration challenges; specialized design limits generalization

D2D-IoT⁶¹

Evaluates lightweight IoT communication protocols under variable network conditions using a testbed

Identifies best protocol for constrained environments; MQTTSN excels in resource efficiency, reliability

Limited to communication layer; does not address physical or MAC layer challenges

CACIE⁶²

Uses data fusion and marine predator algorithm for energy-efficient routing and redundancy reduction

Reduces redundant transmissions, enhances data quality, improves resilience with failure recovery

Complexity in fusion and relay selection; may add overhead for small networks

DD-HAR⁶³

HAR architecture combining IoT devices, edge, and cloud computing for activity recognition with dynamic model

High prediction accuracy (99.23%), low latency on edge devices, supports multiple users dynamically

Dependency on cloud for storage; may require diverse hardware support

PIM-IoT⁶⁴

PIM topologies and multi-tiers, and examine the data flow, and apply PIM structure

each layer to efficiently manage the data processing

Pioneer work, but a multi-tier system under power and latency limitations

MA-LHTO⁶⁵

Multi-stage mixed-integer nonlinear programming model with GPU fragmentation rate and system processing capabilities

Coupled with a deviation-based Lyapunov optimization framework

MA-LHTO multi-agent deep reinforcement learning needs more resources