Electromyography (EMG)-driven musculoskeletal models (EMS) of the trunk are used for estimating lumbosacral joint moments and compressive loads during lifting tasks. These models provide personalized estimates of the parameters using information from many sensors. However, to advance technology from labs to workplaces, there is a need for sensor reduction to improve wearability and applicability. Therefore we introduce an EMG sensor reduction approach based on inverse synergy extrapolation, to reconstruct unmeasured EMG signals for different box-lifting techniques. 12 participants performed an array of tasks (squat, stoop, unilateral twist and bilateral twist) with different weights (0 kg, 7.5 kg and 15 kg). We found that two synergies were sufficient to explain the different lifting tasks (median variance accounted for of 0.91). Building upon this, we used two sensors at optimal subject-specific muscle locations to reconstruct the EMG of four unmeasured channels. Evaluation of the reconstructed and reference EMG showed median coefficients of determination (R2) between 0.70 and 0.86, with median root mean squared errors (RMSE) ranging from 0.02 to 0.04 relative to maximal voluntary contraction. This indicates that our proposed method shows promise for sensor reduction for driving a trunk EMS for ambulatory biomechanical risk assessment in occupational settings and exoskeleton control.