Utility of a ML analytics on real time risk stratification and re-intervention risk prediction on AV access outcomes and cost

Introduction

Vascular access is the lifeline for patients on hemodialysis if peritoneal dialysis is not an option. With 60% prevalent rates of AV fistulas and 15% AV grafts in the United States, it is of paramount importance to preserve the longevity of AV accesses.^1,2 Also, the average survival rates of dialysis dependent patients have been improving over the last 5 years and hence their dialysis access needs extended patency to support uninterrupted optimal dialysis. ³ Numerous factors influence AV access patency.^4–7 Several upstream and downstream attributes contribute to the development of neointimal hyperplasia (NIH). With the lack of clinically available genomic vascular access failure predictors, there is an unmet need for predicting an event and the appropriate approach to mitigate recurrence of the event that could have cost and outcome implications. Loss of functioning AV access results in the need for a tunneled dialysis catheter. The cost and morbidity consequences of dialysis catheters is well known underscoring the importance of preserving functional AV access.

Methods

Local institutional review board review and authorization were obtained for this study. We performed a single center analysis on 200 patients with an AV access (AVF and AVG). Analytics and predictive modeling were performed utilizing a ML model that extracts relevant clinical (access flow, laboratory data and CKD details), access intervention (prior interventions, type & location of lesion, type of balloon used, use of stents etc.) and demographic (age, vintage on dialysis, sex, social determinants, other medical conditions) data in real time and feeds it into validated algorithms to predict risk of reintervention (OptmyCare Inc and Plexus EMR LLC). Two hundred prevalent hemodialysis patients with a AV graft or AV fistula that were functional for at least 3 months were included for this analysis. Demographic, access related, clinical and lab extracts were auto derived from the EMR into the analytics platform. Re-intervention risk prediction models have been validated and tested across 10 years of longitudinal data. Need for re-intervention and use of stent/flow reduction/failed access with TDC placement/new AV access creation were the outcomes analyzed. Risk prediction was available to interventionalist in the validation year to guide treatment options. It was left to physician discretion on the strategy to decrease re-intervention risk which includes inflow banding, covered or bare metal stent utilization, DCB use and surgical revision. The risk for re-intervention was analyzed based on the attributes from the year prior to intervention. Outcomes were retrospectively assessed for a year post intervention after the availability of re-intervention risk prediction. Patients were not prospectively monitored post intervention as the treatments offered were standard of care.

AI technology: Plexus EMR and OptMyCare Inc are licensed Azure based platforms. R and Python software were used to develop and model the ML algorithms. The model was tested across XG Boost, Random Forrest, and multiple regression models. ⁸ Regression factors were developed to assess and test the validity of individual attributes across all the data platforms. Each patient had a real time risk calculator available to the interventionalist to assess risk of reintervention/ year. The analytics do not provide any decision-making outputs suggestive of utilization of one strategy of access treatment over another.

Results

Of the 200 patients included for the analytics, 74% had a AVF and the remaining patients had an AV graft. Since social determinants of health, health disparities and behavioral health have a significant impact on the clinical outcomes and cost in ESRD patients, those factors were included in the analysis for predicting risk. Two patients were transplanted (month 5 and 9 in the validation year), two patients changed to PD (month 6 and 10 of the validation year), three patients died (months 2,8,11). No patients lost their access.

The details of the demographic factors are in Table 1.

OPEN IN VIEWER

Table 1.

Baseline characteristics.

	Fistula	Graft
n	148	52
Age (years)	64±6	66±17
Men	79	24
Women	69	28
Access Vintage (months)	35.4	22.8
Interventions pre-AI	1.8	3.4
Stents placed (pre-AI)	12	10
DM	84	32
HTN	132	46
CAD	56	22
PTH	227.7	291.2
Adequacy	94	95
Hct	31.76	32.23
SDoH (High risk %)	82	91
BH	27	19
Anticoagulants	18	21
Antiplatelets	33	37
Obesity	44	51
Mean systolic BP	137	128
AA	18	22
Hispanic	21	27

Re-interventions: There was a statistically significant reduction in the risk for re-interventions the year post utilization of the non-prescriptive risk prediction tool. The risk reduction ratio is 39% for the need for reinterventions for 1 year post utilization of the tool (Table 2 and 3). There was 40% incremental stent utilization in the post intervention year with almost a doubling in the AV graft group. There was a significant reduction in the need for thrombectomy utilizing the predictive modeling tool (thrombectomy is pooled into reintervention risk) (Table 5 and 6). The risk model does not specifically identify access thrombosis as an output at this time. However, with more modeling thrombosis could be predicted.

OPEN IN VIEWER

Table 2.

No. of interventions Pre and Post AI (1 year).

	Fistula	Graft
Pre-AI	1.8	3.4
Post AI	1.1	2.4
p	<0.01	<0.001

OPEN IN VIEWER

Table 3.

Stent utilization pre and post AI.

	Fistula	Graft
Pre-AI	10	12
Post-AI	15	22

Cost implications

Eighty-two % of the stents used were PTFE coated stents. The cost of treatment was based on claims paid by different payers. The authors had access to vascular access, some dialysis and hospital physician component claims, the rest of Part A, B and D claims were not available hence cumulative cost on each member cannot be assessed. Only 11% of the patients changed their carrier between the assessment and post AI tool year and all of them enrolled from a traditional medicare to a medicare advantage plan. The total paid claims for AV access interventions were compared between the members before and after re-intervention risk tool utilization. Despite the increase in stent use, there was a 200k cumulative cost reduction in the observation year which correlates with the decreased need for reinterventions. There is an approximately $1000 cost reduction PMPY (per member per year) in the post AI tool implementation year. The risk prediction tool was able to accurately identify the members at the highest risk for reintervention in both the AVF and AVG groups. The cost implications were calculated accounting for the members who dropped out of the validation period based on how long they were active during the validation year of the study. (Table 4).

OPEN IN VIEWER

Table 4.

Cost implication.

Pre-AI	$712,609.00
Post AI	$512,172.00

OPEN IN VIEWER

Table 5.

Reintervention risk predictions in %.

	Fistula	Graft
>3	4	12
3	5	7
2	5	11
1	6	18
0	80	52

OPEN IN VIEWER

Table 6.

Thrombectomies Pre and post AI.

	Fistula	Graft
Pre-AI	6	56
Post-AI	0	34

Discussion

The estimated number of patients with ESRD in the next 5 years could be over 1 M. ⁹ Even with the assumptions of a significant increase in peritoneal dialysis there would still be more fistulas and grafts than what we have today in the near future.^10,11 There is a looming threat of decreasing healthcare workforce and spiraling cost of care mandating the need for identifying emerging risk in every disease.^12–14 About 5%–10% of members contribute to more than 50% of cost in any cohort of patients. It is critical to identify these members at risk of emerging risk for worsening disease and outcomes to bend the disease and cost curve.^15,16

The first steps in decreasing cost in AV access care is performing a vast majority of access creation and salvage procedures in an outpatient setting.^17-19 Utilizing a ML algorithm could help reduce cost of care by 74% by directing AV access procedures to specialized vascular access outpatient clinics (Balamuthusamy et al. Oral presentation at ASDIN 2021). Further improvement in outcomes and reductions in cost could be facilitated by utilizing ML analytics that utilizes a variety of attributes that takes clinical, social determinants, behavioral health, health inequity and access related parameters into consideration. We utilized data extraction utilizing Plexus EMR which has built-in ML tools for re-intervention risk prediction. All the non-access related data is processed through OptMyCare’s ML platform to analyze the impact of SDOH, MH and health inequity. The models are trained to intake recurrent data feeds and adapted to scavenge irrelevant attributes. The model functions with 10%–15% supervision and the rest are unsupervised learning.

Access monitoring tools like wearables, transonics, vascalert and needle reversal are used for monitoring flow as a single parameter which could be correlated with a functional stenosis. However, flow changes are dependent on numerous factors like changes in MAP and CV status.^19,20 Reacting to flow changes could be burdensome and also not all stenotic lesions need to be addressed as some at the inflow could be protective from downstream adverse cardiovascular events. ²¹ Hence there is a need for a more comprehensive methodology to identify patients at risk for reinterventions.

Though the data presented only includes 200 patients, it needs to be underscored that several thousand procedures and data points over 10 years were utilized to develop the models. The accuracy of the predictive analytics has been validated and averages greater than 86% across all chronic conditions. This is the first validation of a ML model in AV access literature that predicts emerging risk of reinterventions with improved access outcomes and decreased cost. The next steps would be to perform a case control trial utilizing the model versus standard of care to assess the efficacy across a large volume of patients.

Most ESRD patients are part of a value-based care system irrespective of their insurance status. Value based care systems hold the risk bearing entities accountable on both clinical outcomes and cumulative cost. Utilizing precision analytic systems backed by ML models could be the future of optimizing the overall health ecosystem that mandates improving patient reported outcomes, optimizing healthcare workforce utilization, improving hard clinical outcomes and decreasing overall cost of care.

In conclusion, we present the first ever prospective observational study that validates the utility of a ML risk prediction model that demonstrates improved outcomes and reduces access related cost. This is not a randomized trial and also is not a multicenter project, hence it needs further validation across multiple providers and several clinical sites.