We want to predict body weight while lying in bed for an elderly patient who is unable to move by himself/herself. To this end, we have implemented a prototype system that estimates the body weight of a person lying on a smart mat in nonrestraint and unconsciousness conditions. A total of 128 FSR (force sensing resistor) sensors were placed in a 16 × 8-grid structure on the smart mat. We formulated three methods based on the features to be applied: segmentation, average cumulative sum of pressure, and serialization. All the proposed methods were implemented with four different machine-learning models: regression, deep neural network (DNN), convolutional neural network (CNN), and random forest. We compared their performance using MAE and RMSE as evaluation criteria. From the experimental results, we chose the serialization method with the DNN model as the best model. Despite the limitations of the presence of dead space due to the wide spacing between the sensors and the small dataset, the MAE and the RMSE of the body weight prediction of the proposed method was 4.608 and 5.796, respectively. That is, it showed an average error of ±4.6 kg for the average weight of 72.9 kg.

中文翻译：

---

预测在无约束和无意识条件下躺在智能垫上的人的体重。


我们想要预测无法自行移动的老年患者躺在床上时的体重。为此，我们实现了一个原型系统，可以估计在非约束和无意识条件下躺在智能垫上的人的体重。智能垫上的 16 × 8 网格结构中总共放置了 128 个 FSR（力敏电阻）传感器。我们根据要应用的特征制定了三种方法：分段、平均压力累积和和序列化。所有提出的方法都是用四种不同的机器学习模型实现的：回归、深度神经网络（DNN）、卷积神经网络（CNN）和随机森林。我们使用 MAE 和 RMSE 作为评估标准来比较他们的表现。从实验结果来看，我们选择了以DNN模型为最佳模型的序列化方法。尽管由于传感器之间的间距较宽和数据集较小而存在死腔的限制，但该方法体重预测的 MAE 和 RMSE 分别为 4.608 和 5.796。也就是说，对于 72.9 公斤的平均重量，其平均误差为 ±4.6 公斤。