CHAPTER 2: A PRECISION MEDICINE APPROACH TO DEVELOP AND IN-
2.4 Discussion
In this paper, we investigated optimal treatment recommendations for older and overweight or obese individuals with KOA using precision medicine techniques and machine learning tools applied to data obtained from the IDEA trial. The individual treatment decisions obtained from our precision medicine approach are data-driven (requiring no strong assumptions), reproducible (with careful reporting of the analysis process) (Kosorok and Laber, 2018), and generalizable and extendable to other clinical settings (because of rich heterogeneity in the clinical input data).
The results of the optimal ZOM, where everyone would be assigned to a single intervention, match with those from the published IDEA trial (Messier et al., 2009, 2013). The assignment of patients to the D+E intervention would be expected to result in the optimal improvement in the majority of patients in the clinical outcomes of weight loss since baseline, WOMAC pain, function, and stiffness scores, as well as PCS and so should remain the recommendation of choice. In individuals where the primary goal is to reduce systemic inflammation as measured by plasma IL-6 levels and/or reduce the knee compressive force, D alone would be the treatment of choice.
The optimal treatment rules of the optimal PMMs suggested that not everyone benefits from D+E even though patients are expected to be assigned to this group based on the ZOM. Further improvements in weight loss could be obtained in certain patients selected by measures of high baseline weight (over 109.35 kg) or low waist circumference (90.25 cm or less) accompanied by lack of a previous heart attack that would result in assigning them to D rather than D+E. This would only be a consideration if weight loss alone was more important to the patient than the level of improvement in pain and function. We can only speculate why people of higher weight with lower waist circumference and no history of heart attack would benefit more from D than D+E. First, it is likely that following the suggested exercise program may be more difficult for patients with a height weight. Second, higher weight with lower waist circumference could be
seen in individuals who have more peripheral adiposity rather than central adiposity. In these cases, D could be more effective in losing weight. The finding that our results were modified by a history of a heart attack may be that the cardiac status of these individuals encourages optimal compliance and improves more with the combined D+E than D alone and this allows for greater activity levels resulting in greater weight loss.
The finding that the IL-6 outcome improves more with D than D+E in certain individuals is not easily explained. We noted that individuals with high baseline IL-6 levels (i.e. above 4.5 pg/mL) or those with low baseline function scores (12.5 or less in a range of 0 to 68) reduced their IL-6 more from diet only. Individuals whose IL-6 was not high but have poorer function are recommended to receive both diet and exercise. The decrease in IL-6 suggests less systemic inflammation but there is no solid evidence to suggest that exercise would modulate the reduction in IL-6 that occurs with dietary weight loss. Because all three groups received an intervention, the significant differences in outcomes noted among the groups at 18 months would be unlikely to be due to regression to the mean. Our findings that specific subgroups of individuals received more benefits from specific interventions argues against the premise that response was simply due to patient perception rather than to the intervention itself.
As for the multiple outcomes, comparison between Table 2.4 and Table 2.5 suggested that our minimax rule together with the coarse-to-fine grid search for parameter optimization can be a useful way to incorporate multiple outcomes, and combining correlated outcomes has the potential for bringing more benefits to patients than single outcomes. However, uncorrelated outcomes do not benefit from the composite outcome.
2.4.1 Limitations
Potential limitations of this study include the following. First, we were not able to use the information of about 100 of the trial participants due to missing outcome data. Although larger sample size might lead to higher power, our two input datasets remain representative of the overall data as Table 2.3 shows. Secondly, the analyses did not include intermediate follow-up data at 6 months. Although longitudinal analysis methods could be applied to the IDEA data, we
were more interested in the final improvements of each outcome between the start and the end of the trial and less on the intermediate progress. It is also unlikely for adding one more time point shortly after the trial would be influential as we expect it takes time for the interventions to take effect. Thirdly, there were some covariates with a large proportion of missing data excluded from the analysis. The majority of these were measures that would not be routinely collected in the clinical setting such as full-length lower extremity radiographs for alignment, computed tomography for abdominal and thigh fat, knee MRI, and isokinetic strength testing. Finally, our results are from a single clinical trial of patients with mild-to-moderate symptomatic KOA (Kellgren-Lawrence scores of 2-3) (Messier et al., 2013) and may not be generalizable to populations with more severe KOA.
2.5 Supplementary Materials