Machine learning for predicitng the risk of osteoporosis from patient attributes, health and lifestyle history

Abstract

The most widely-used method for diagnosis of osteoporosis is to determine bone mineral density (BMD) by bone densitometry. At present mass screening is not, on the basis of resource constraints, considered a option. This project investigates if artificial neural networks (ANN s) or Baysian networks (BNs), using the health and lifestyle history of a patient, (risk factors - used as a generic term for inputs) may be used to develop a preliminary screening system to determine in a patient is at particular risk from osteoporosis and hence in need of a scan. Two databases have been used, one containing 486 records (29 risk factors) of patients examined with a G E Linear Peripheral Densitometer (PIXI) and the other with 4,980 records (33 risk factors) of patients examined with dual energy X ray absorptiometry (DEXA). BNs tend to out-perform AN s particularly where smaller learning sets are involved. The best result was 84% accuracy (sensitivity 0.89 and specificity 0.80) with PIXI and a BN. I general, however, with ANNs the sensitivity achieved with PIXI and DEXA was 0.65 and 0.80 respectively and the corresponding values with BNs were 0.72 and 0.81. The diagnostic performance with ANNs could be achieved with fewer risk factors (PDQ from 29 to 4 and DEXA from 33 to 5) but with BNs a reduction in performance accompanied a reduction in the number of risk factors. l The results also indicate: 0 For Positive patients, the more severely affected by the disease the more accurately they are diagnosed . 0 The lack of continuous values in the DEXA data results in a poor diagnosis of Negative patients. 0 Classifications based on BMD predictions and pattern recognition give similar results. 0 Reasoning with BNs can provide an indication of how a particular risk factor state contributes to a patient`s risk from osteoporosis.