Personalized treatment options for chronic diseases using precision cohort analytics

Personalized treatment options for chronic diseases using precision cohort analytics

Study designThe PCTO workflow was applied to de-identified EHR data from a large health care provider in eastern Massachusetts, containing over 20 years of longitudinal data for approximately 2.5 million patients, for three disease conditions: HTN, T2DM, and HL. The Western Institutional Review Board (WIRB) had reviewed the study and gave it an exemption as a human subjects study. The decision-point criteria and characteristics for the three disease conditions are described in Table 2. The table lists the clinical treatment guidelines referenced, the start and end dates of the study period used for extracting data from the EHR, the decision-point criteria, the specifications for the baseline period, baseline variables, treatment options, follow-up period, and outcome. The clinical treatment guidelines were selected based on what was currently in use at the pilot health system. The study period was selected to include data covered by the time of the JNC7 publication to the present (for HTN), and the most recent 10 years of data (for T2DM and HL), so only recent and relevant treatments and practice patterns are included. The DP criteria, baseline period, follow-up period, and outcome specifications were based on clinic expert opinion. The baseline variables were derived using the process described in “Data extraction”. The treatment options were manually extracted from the guidelines, so only clinically acceptable treatments were included. Finally, the table shows the number of unique patients selected, and the number of DPs extracted based on these criteria. After the DP extraction, 50,000 DPs were selected at random to form the training data set. The remaining DPs were used as the scoring data set, following the process described in “Data extraction”. A disease-specific LSML similarity measure was trained using the training data set as described in “Similarity model training”. To perform runtime analysis, a leave-one-out cross-validation (LOOCV) approach51 was adopted. Each DP in the scoring data set was selected, in turn, to represent the individual patient’s office visit. That DP was then removed from the scoring data set, and the precision cohort identification (from “Precision cohort identification”) and treatment options analysis (from “Treatment Options Analysis”) were performed using the selected DP of interest and the updated scoring data set. Finally, the personalized treatment option outputs were aggregated and analyzed. The following sections present results from analyzing the data and output for individual DPs, groups of DPs, and all DPs.Table 2 Decision point criteria and characteristics for the three disease conditions.In terms of computational cost, the extraction of the data to create the training and scoring data sets typically took under one processor-hour. Creating a new scoring model took between 4 to 8 processor-hours, and the dynamic generation of the precision cohort and personalized treatment options report for an individual patient of interest took a few processor-seconds. The entire LOOCV analysis typically took 3000 processor-hours (since all DPs had to be processed). However, since the runtime was parallelizable, the analysis was performed on a cluster of approximately 160 processors, thus providing results within a day.Filter variable cohortsThe largest observed cohorts based on the “exact match” filter variables for HTN, T2DM, and HL are shown in Supplementary Figs. S5, S6, and S7, online, respectively. Since the filter variables capture the active treatments and key patient characteristics used to inform treatment decisions (based on the clinical guidelines), these cohorts represent the most commonly observed active treatment combinations at the DPs. Although there is a large number of cohorts (32,656 for HTN, 1867 for T2DM, and 253 for HL), it is much less than all the possible combinations (290 for HTN, 248 for T2DM, and 214 for HL); moreover the cumulative coverage curve indicates that most DPs belong to only a small number of cohorts (40% in the top 10 cohorts for HTN, 62% for T2DM, and 91% for HL) and that there is a long tail (70% of the groups have less than five DPs for HTN, 60% for T2DM, and 34% for HL). These observations are consistent with other published findings where the number of unique treatment pathways for HTN and T2D is large and diverse52. Unsurprisingly, the largest cohorts are those that are not currently on any treatment, followed by cohorts that are only on a single treatment (e.g., Lisinopril for HTN, Metformin for T2DM, and Simvastatin for HL). For HTN and T2D, the size of the “no treatment” cohorts may be increased because the selection criteria (Table 2) can include DPs where the patient does not have an official disease diagnosis (i.e., based only on abnormal lab test measurements). It is also possible that some diagnosed patients are only on lifestyle (e.g., diet and exercise) modification treatments that are not explicitly tracked in the analysis. Although an explicit disease diagnosis is required for HL, there is variability in the initiation of pharmacologic treatment based on LDL levels, which results in more “no treatment” DPs (especially for younger patients and those with borderline LDL levels)53. In addition, similar to HTN and T2D, some HL patients may only be on non-medication lifestyle modification treatments.Global treatment optionsThe largest observed global treatment option decisions across all DPs for HTN, T2DM, and HL are shown in Supplementary Figs. S8, S9, and S10 online, respectively. Across all three conditions, the most common treatment decision is to “stay the current course” and make no change to the current medication-treatment ingredient (81% for HTN, 85% for T2DM, and 94% for HL). Note that since the dosage was not considered, treatment decisions that only changed the dosage appeared as refills and were considered “no ingredient change.” For HTN and T2DM, some no-change treatments reflect no active pharmacologic treatments for undiagnosed patients. Even for undiagnosed patients who have measurement-proven conditions, however, “clinical inertia” occurs when a PCP does not treat a condition that is not under control. Due to the large number of DPs without any active treatments, the next most common set of treatment option decisions are the start of first-line medications: Lisinopril (ACEI) and Hydrochlorothiazide (thiazide diuretic) for HTN, Metformin (biguanides) for T2DM, and Simvastatin and Atorvastatin (statin) for HL. Overall, there are many treatment option decisions (3966 for HTN, 883 for T2DM, and 223 for HL). However, many are unique and have only a single DP (60% for HTN, 53% for T2DM, and 29% for HL), again consistent with prior findings on the uniqueness of treatment pathways52. In terms of treatment decisions that switch ingredients within the same class, it is interesting to note that, for T2DM, the most common are Sulfonylurea changes: stopping Glyburide and starting Glipizide or Glimepiride. For HL, the most common are Statin changes: stopping Simvastatin and starting Atorvastatin, Pravastatin, or Rosuvastatin.Personalized treatment optionsFigure 4 illustrates the patient-level personalized treatment options and associated outcomes, resulting from applying the PCTO analysis to a single DP for an individual patient. In this case, the patient has an HTN diagnosis, is over age 65, and is on Lisinopril (ACEI) at the time of the DP. In addition to identifying treatment options that have statistically significantly better outcomes (dark green pathways) than the baseline “no change” treatment option (grey pathway), the analysis also lists treatment options that have significantly worse outcomes (dark red pathways). Options in light green or light red have better or worse outcomes, respectively, but are not statistically significant. The complexity of each treatment option, estimated as the number of unique medications, is also available.DPs from different filter variable cohorts are expected to have different personalized treatment options. For example, Fig. 5A shows the treatment options for a DP, in a group that is over age 60, and on Hydrochlorothiazide (Thiazide), while Fig. 5B shows the treatment options for a DP, in a group that is over age 60, and on Lisinopril (ACEI). The treatment options associated with significantly better-than-the-baseline outcomes (dark green pathways), for these two DPs, are different: in Fig. 5A, it is adding Lisinopril (ACEI), while in Fig. 5B, it is adding either Amlodipine (CCB) or a Thiazide (Hydrochlorothiazide or Chlorthalidone). However, patient DPs that belong to the same filter variable cohort can also have different personalized treatment options. Figure 5C shows the personalized treatment options for a patient DP that belongs to the same group as Fig. 5B [over age 60 and on Lisinopril (ACEI)], but the effective treatment options here are different: only adding Hydrochlorothiazide passes the selection criteria. Since the precision cohort is created dynamically for each patient DP, based on its specific characteristics, the composition of the precision cohort can differ, resulting in different personalized treatment options.Figure 5Personalized treatment options for three different HTN patient decision points. (A) The patient decision point belongs to a filter variable group that is over age 60 and on Hydrochlorothiazide (Thiazide). (B, C) The patient decision point belongs to a filter variable group that is over age 60 and on Lisinopril (ACEI). Note that the treatment options associated with significantly better control (dark green pathways) across the three patient decision points are different.To assess the applicability and potential impact of the PCTO method, and to better understand its behavioral characteristics, all DPs in the scoring data set were processed using the LOOCV strategy described in “Results”; the results are summarized in Supplementary Figs. S11, S12, and S13 online (for HTN, T2DM, and HL, respectively). Each figure shows the number of DPs, grouped by filter variable cohorts (rows), and the percentage of DPs that have personalized treatment options with statistically significantly better outcome (green), with fewer medications (purple), with the same number medications (yellow), and with more medications (red). As expected, filter variable cohorts with no active medication treatments (e.g., “All guideline variables false” for all conditions, “age_60_or_above” for HTN, “T2D” for HL) have the largest percentage with significantly better treatment options. For HTN, it is interesting to note that for the African American (“black”) group on no active HTN medications, only 70% of the DPs have treatment options with statistically significantly better outcomes compared to taking no HTN medications. Similar behavior is seen in the diabetes (“dm”) group on no active HTN medications where only 78% of the DPs have treatment options with statistically significantly better outcomes. This implies that it may be more challenging to control HTN for patients that are black or have diabetes and is consistent with prior findings54,55. With a few exceptions, filter variable cohorts with one active medication treatment (e.g., “ACEI_Lisinopril” or “Thiazide_Hydrochlorothiazide” for HTN, “Biguanides_Metformin” and “Sulfonylurea_Glyburide” for T2DM, and “Statin_Simvastatin” and “Statin_Pravastatin” for HL) have the next largest percentage with significantly better treatment options. The notable exceptions are “Beta_Blockers_Metoprolol” for HTN, “Insulin_Glargine” for T2DM, and “Statin_Atorvastatin” and “Statin_Rosuvastatin” for HL. Indeed, beta blockers for HTN and insulin for T2DM are not typically first-line treatments, and patients on these medications likely have already tried (unsuccessfully) other medications before settling on these. For HL, atorvastatin and rosuvastatin are the most effective statins available56 and would be difficult to improve upon. Recognizing that successful disease management should not only consider disease control, but other factors such as treatment complexity and cost57, information about the number of medications relative to the baseline (fewer, same, more) for each treatment option is also computed. Across the three diseases, many filter variable cohorts that are on multiple medications, have treatment options with fewer or the same number of medications with comparable outcomes (percent controlled). Treatment options with more medications are almost always possible. The impact and utility of PCTO are dependent on the specific filter variable cohort, but overall, 75%, 74%, and 85% of all the DPs for HTN, T2DM, and HL, respectively, have at least one significantly better treatment option identified.

Via Source link