Power of algorithm: Identifying high-risk non-adherent patient groups
- Non-adherence to medication is a major challenge to treatment success and a significant burden on the healthcare system1-3.
- The existing interventions are based on retrospective measurements of non-adherence in patients already undergoing treatment and are therefore limited in improving future medication adherence4.
- Predictive analytics can help in identifying potentially high-risk non-adherent patient groups at the beginning of the therapy to help provide tailored interventions5.
Medication adherence is essential for successful treatment, especially for long-term chronic diseases. Adherence simply means taking the medicine or performing intervention as prescribed by doctors. Non-adherence to therapy can make any treatment ineffective, leading to suboptimal health results. This may further escalate to be the cause of increased morbidity and mortality1-3. Non-adherence is also a significant economic burden and leads to considerable medication wastage every year3. Some of the key reasons for non-adherence to the drugs as prescribed by the clinicians are:
- Asymptomatic nature of the disease
- Longterm usage of medicines
- Adverse effects of the medication
- Complex dosing regimen
- Multiple morbidities
- Psychological issues
- Economic constraints
- Lack of social support
- Poor patient motivation
However, it is not easy to gauge the presence of these factors in a patient while prescribing the medication. Most of the assessment for non-adherence happens retrospectively, either post-therapy or sometimes while continuing the treatment during routine follow-ups. This is a major limitation in addressing non-adherence to the prescribed therapy. Therefore, we need measures that can distinguish patients at high risk of developing non-adherent behavior from others at the time of therapy initiation4.
Conventional methods of measuring adherence and scope of predictive analytics:
Broadly, methods for measuring adherence can be classified into two major types: direct and indirect methods. Direct assessment methods like measurement of drug/metabolite levels in body fluids etc. are ideal but labor and cost-intensive. Adherence is assessed mostly via Indirect methods like pill counting, self-reporting by patients, using electronic databases, and surveys or questionnaires, since they are easier to employ3. For more on this, see our other article: Measuring adherence – an “Achilles heel” in medication adherence.
As mentioned above, these methods are used retrospectively either post the treatment or while patients are well into the treatment. By then, the harm is done in terms of reducing the effectiveness of the therapy resulting in undesired health consequences. Gaining insight into patients’ future non-adherence behavior using predictive analytics at the therapy initiation stage will be a game-changer in curbing sub-optimal medication usage with the potential to enhance treatment success as reported in clinical trials6 .
What is predictive analytics?
As the name suggests, predictive analytics is an algorithm-based technique to envisage non-adherence behavior in patients. Even unintentional non-adherence has some intent behind it, and predictive analysis may help to recognize it. Such analytics uses cutting-edge technology such as artificial intelligence, machine learning, and real-time measurement. It can help not just in predicting non-adherence, but also its underlying causes to enable a targeted, tailored, and wholesome intervention1,6,7.
The two essential components of predictive analytics are data and predictive models7. The predictive models are based on advanced computational algorithms that use the available past data to make an educated prediction about future outcomes. Therefore, the quality of data is the most important determinant of the efficiency of predictive models and also its major limitation. Various computational tools are used in prediction analytics, a few are listed here:
- Machine learning8
- Ensemble learning and Deep learning9
- Logistic regression8
- Classification and regression tree (CART)8
- Non-Adherence Tree Analysis (NATA)6
- SPUR (Social Psychological Usage Rational)10
A few of these that have been successfully used in predicting non-adherence behavior are discussed below in some detail.
Machine Learning: Machine learning algorithms include bayesian network, neural network, and support vector machines to make predictive models. It was successfully used to prevent non-adherence risks in patients with diabetes11. It was also used to develop a digital tool for identifying non-adherent patients at the early stages of Crohn’s Disease. Interestingly, machine learning-based models revealed anxiety, depression, and medication beliefs as important risk factors for non-adherence8.
Deep learning: It is a subfield of machine learning and includes deeply layered neural networks that behave like the brain in terms of computations. Ensemble Learning mixes various algorithms in making novel and unique models. One use of deep learning model captured injection administration and disposal data in real-time and generated a large scale data-set. This was used to assess adherence to the therapy and importantly, to identify patient characteristics prone to non-adherence. The data set could be used to predict patients at risk of non-adherence in the future9.
NATA: Non-Adherence Tree Analysis uses Fault Tree Analysis (FTA) and Monte Carlo simulation as their basis for predicting adherence. FTA is an analysis technique to identify the root cause of non-adherence behavior in patients. Monte Carlo simulation is a mathematical algorithm that uses repeated random sampling and predicts the behavior of a system. The non-adherence factors are used to prepare the Non-adherence Tree (NAT) that is analyzed by Monte Carlo simulation to predict behavior based on these factors. This model was recently used to predict adherence to COVID-19 treatment. The study found patient-related factors like concern for side effects, forgetfulness, and lack of symptoms as the key contributor to non-adherence6.
SPUR: The SPUR analysis is a questionnaire tool with foundations in behavioral theories like the theory of planned behavior and the health belief model10. SPUR hypothesize that patients’ action towards a particular treatment is based on attributes like their behavioral attitude, social and subjective norms, and their perception of control. Specifically, SPUR stands for the following four parameters:
- Social: The patient’s belief about social norms based on the social factor
- Psychological: Psychological aspect includes three main frameworks – the concept of the self, reactance, and discounting of future values
- Usage: Usage deals with attributes like patients’ financial concerns, own usage capability, and forgetfulness
- Rational: Rational factors are associated with expected outcome benefits and perceived threats10.
SPUR analysis uses cognitive interviews with patients to get relevance and validation of the parameters and utilizes this along with behavioral theories. It is majorly used in chronic diseases with patients undergoing long-term treatment like diabetes12. SPUR questionnaire and algorithm validation have produced insightful data for non-adherent behavior and the underlying causes13. It can identify patient-specific attributes for non-adherence to formulate individualized and tailored solutions for enhancing adherence14.
Predictive analytics tools –strengths and limitations
Advanced predictive algorithm tools discussed above are effective in predicting non-adherence behavior and the associated factors6,14. They could help in providing tailored interventions for meeting patients’ unique needs leading to better adherence and improved treatment efficacy5. Increasing medication adherence with the help of predictive techniques will also help in reducing healthcare costs and medication wastage. These big data analytics tools can give a holistic view of patients that are most likely to be non-adherent based on their attributes, beliefs, and self-reported information. This will assist clinicians in identifying the predictors of non-adherence and provide appropriate interventions at an early stage in the treatment2.
However, since most predictive models, except SPUR, use data from a specific population, they cannot be generalized to a larger patient pool. These techniques cannot assess all predictors of non-adherence. Younger, elderly, or specific population data which are not fed in records or registries cannot be evaluated2. Predictions from models using self-reported data in the form of questionnaires or surveys are less reliable due to the quality of the founding data. Real-world studies will be required to validate the accuracy of these tools. A more comprehensive as well as extensive data set covering different populations with different age groups and various factors associated with non-adherence is required for these techniques to be more precise and validated11.
Medication adherence is the single most important measure that can improve treatment success rates across therapies. Predictive algorithms can predict the non-adherence factors and future non-adherent patients but are limited by the quality and quantity of past data used as the basis for making these predictions. With the implementation of sizeable multifaceted and diverse data sets, these tools have the potential to be a game changer in the healthcare landscape.
1. Gadkari AS, McHorney CA. Unintentional non-adherence to chronic prescription medications: how unintentional is it really? BMC Health Serv Res. Jun 14 2012;12:98. doi:10.1186/1472-6963-12-98
2. Zullig LL, Jazowski SA, Wang TY, et al. Novel application of approaches to predicting medication adherence using medical claims data. Health Serv Res. Dec 2019;54(6):1255-1262. doi:10.1111/1475-6773.13200
3. Anghel LA, Farcas AM, Oprean RN. An overview of the common methods used to measure treatment adherence. Med Pharm Rep. Apr 2019;92(2):117-122. doi:10.15386/mpr-1201
4. Kardas P, Lewek P, Matyjaszczyk M. Determinants of patient adherence: a review of systematic reviews. Front Pharmacol. 2013;4:91. doi:10.3389/fphar.2013.00091
5. Zullig LL, Blalock DV, Dougherty S, et al. The new landscape of medication adherence improvement: where population health science meets precision medicine. Patient Prefer Adherence. 2018;12:1225-1230. doi:10.2147/PPA.S165404
6. Edifor EE, Brown R, Smith P, Kossik R. Non-Adherence Tree Analysis (NATA)-An adherence improvement framework: A COVID-19 case study. PLoS One. 2021;16(2):e0247109. doi:10.1371/journal.pone.0247109
7. Koesmahargyo V, Abbas A, Zhang L, et al. Accuracy of machine learning-based prediction of medication adherence in clinical research. Psychiatry Res. Dec 2020;294:113558. doi:10.1016/j.psychres.2020.113558
8. Wang L, Fan R, Zhang C, et al. Applying Machine Learning Models to Predict Medication Nonadherence in Crohn’s Disease Maintenance Therapy. Patient Prefer Adherence. 2020;14:917-926. doi:10.2147/PPA.S253732
9. Gu Y, Zalkikar A, Liu M, et al. Predicting medication adherence using ensemble learning and deep learning models with large scale healthcare data. Sci Rep. Sep 23 2021;11(1):18961. doi:10.1038/s41598-021-98387-w
10. Dolgin K. The SPUR Model: A Framework for Considering Patient Behavior. Patient Prefer Adherence. 2020;14:97-105. doi:10.2147/PPA.S237778
11. Wu XW, Yang HB, Yuan R, Long EW, Tong RS. Predictive models of medication non-adherence risks of patients with T2D based on multiple machine learning algorithms. BMJ Open Diabetes Res Care. Mar 2020;8(1)doi:10.1136/bmjdrc-2019-001055
12. de Bock E, Dolgin K, Arnould B, Hubert G, Lee A, Piette JD. The SPUR adherence profiling tool: preliminary results of algorithm development. Curr Med Res Opin. Feb 2022;38(2):171-179. doi:10.1080/03007995.2021.2010437
13. de Bock E, Dolgin K, Kombargi L, et al. Finalization and Validation of Questionnaire and Algorithm of SPUR, a New Adherence Profiling Tool. Patient Prefer Adherence. 2022;16:1213-1231. doi:10.2147/PPA.S354705
14. Tugaut B, Shah S, Dolgin K, et al. Development of the SPUR tool: a profiling instrument for patient treatment behavior. J Patient Rep Outcomes. Jun 6 2022;6(1):61. doi:10.1186/s41687-022-00470-x