Application of a Neural Network Classifier to Radiofrequency-Based Osteopenia/Osteoporosis Screening

There is an unmet need for quick, physically small, and cost-effective office-based techniques that can measure bone properties without the use of ionizing radiation. The present study reports application of a neural network classifier to the processing of previously collected data on very low power radiofrequency propagation through the wrist with the goal to detect osteoporotic/osteopenic conditions. Our approach categorizes the data obtained for two dichotomic groups. Group 1 included 27 osteoporotic/osteopenic subjects with low BMD (DXA T score below - 1) measured within one year. Group 2 included 40 healthy and mostly young subjects without major clinical risk factors such as (family) history of bone fracture. We process the complex radiofrequency spectrum from 30 kHz to 2 GHz. Instead of averaging data for both wrists, we are processing them independently along with the wrist circumference and then combine the results, which greatly increases the sensitivity. Measurements along with data processing require less than 1 min. For the two dichotomic groups identified above, the neural network classifier of the radiofrequency spectrum reports a sensitivity of 83% and a specificity of 94%. These results are obtained without inclusion of any additional clinical risk factors. They justify that the radio transmission data are usable on their own as a predictor of bone density. This approach has the potential for screening patients at risk for fragility fractures in the office, given the ease of implementation, small device size, and low costs associated with both the technique and the equipment.

目前临床仍缺乏快速、便携且成本低廉的非电离辐射式骨特性检测门诊技术。本研究将神经网络分类器（neural network classifier）应用于此前采集的极低功率射频信号经手腕传播的数据集处理，旨在检测骨质疏松/骨量减少状态。本研究针对两组二分法分组的采集数据开展分类：第一组包含27名1年内经检测确诊为低骨密度（Bone Mineral Density, BMD，由双能X线吸收法（DXA）测得的T值低于-1）的骨质疏松/骨量减少受试者；第二组包含40名健康且以年轻人群为主的受试者，无骨折（含家族性骨折史）等主要临床骨病风险因素。本研究处理30 kHz至2 GHz频段的复杂射频频谱，相较于对双侧手腕数据取平均的常规方法，我们分别处理单侧手腕数据并结合手腕周长参数，随后整合分析结果，这一策略大幅提升了检测灵敏度。整个检测流程与数据处理耗时不足1分钟。针对上述两组二分法分组，基于射频频谱的神经网络分类器实现了83%的灵敏度与94%的特异度，且该结果无需引入任何额外临床风险因素即可获得。上述结果证实，仅依靠射频传输数据即可作为骨密度的有效预测指标。鉴于该方法实施简便、设备体积小巧且技术与设备成本低廉，其有望在门诊场景中实现脆性骨折风险人群的筛查。