Predicting Patient Adopters Through Deep Learning

Key Takeaways:

• Deep learning automatically processes complex data sets with minimal feature engineering. As a result, health care professionals match patients to the proper treatments much faster than they could with previous machine learning models.
• Pros and Cons of Different Predictor Models: What predictor models exist? How can analysts choose the right one based on a specific business objective and the right interpretability-accuracy balance?
• Use Case: Using deep learning, Axtria, Inc. predicted the likelihood that doctors would prescribe specific patients specific drugs with over 80% accuracy — a ~10% increase compared to previous methods.

In pharma, especially in specialized areas like oncology, precision medicine requires an understanding of accurate patient cohorts and their required treatments with drugs to drive better outcomes. However, different patient types make it difficult to connect the right patient to the right medicines. In such cases, deep learning (DL) automates patient cohorts right treatments for the right patients at the right time without manual feature engineering.¹

By enhancing the ability to dive into complex data sources like highly-dimensional patient data, DL uncovers newer, accurate insights to aid decision-making that can positively impact patient outcomes.

By picking patterns from the complex data autonomously with minimal human supervision, DL models can:

• Make better and accurate predictions.
• Help interpret complex unstructured data.
• Produce highly-actionable outcomes by bringing out deep, specific insights.

Currently, professionals utilize DL for a variety of healthcare applications, including diagnosing diabetes through analysis of retina images, detection of skin cancer by analyzing images of skin, and several recent analyses of electronic health records (EHR) data.

Real-World AI Applications

Recently, Google AI demonstrated the impact of DL application.² They predicted the likelihood of a patient’s readmission with high accuracy by applying DL on EHR data. The models used thousands of potential predictor variables in patient records. In turn, the model indicated specific events in the patient’s history to understand why they were likely to be readmitted.

Their results were statistically significant. On a scale where 1.00 is perfect, and .50 is no better than random chance, the scores outperformed traditional logistic regression as follows:

• 0.86 in predicting if patients will stay long in the hospital versus 0.76 in traditional methods.
• 0.95 in predicting inpatient mortality, versus 0.86 in traditional methods.
• 0.77 in predicting unexpected readmission after patients are discharged, versus 0.70 in traditional methods.

Collectively, DL outperformed traditional methods by nearly 10%. Overall, Google demonstrated the high-efficiency of DL in extracting insights from high volumes of data minus manual featuring.

DL Approaches to Predicting Patient Adopters

Generally, two broad methods exist to determine patient adopters. Classification models yield insights for entire patient populations. On the other hand, more specific models yield insights specific to individual patient cohorts.

Classification models are developed to anticipate adopters for the entire patient populations. Because this method accounts for all patients within a single data set, there is a risk for generic insights. Ultimately, it provides a unique set of insights for several different patient types under the same umbrella.

Patients, for specialized areas like oncology, require highly-specific treatments. Since each patient is highly unique, models that develop insights into different patient cohorts are ideal. In these cases, the model groups patients with similar characteristics in terms of their treatment patterns.

Therefore, unsupervised DL is beneficial for automatically generating specific patient cohorts to predict adopters. To do so, these models extract features from among thousands of input variables, breaking down complexities of multi-dimensional data sets without manual intervention. The steps to create these patient groups are as follows:

1. Identify patient cohorts using unsupervised techniques such as Restricted Boltzmann Machines and auto-encoders.
2. Develop individual models to estimate the likelihood of patient adoption of the drug of interest. The model can be interpreted to determine which patient characteristics (i.e., co-morbidities, demographics, past treatment patterns, attending providers and HCPs, etc.) will drive the adoption.

Pros and Cons of Various Predictor Models

Choosing a model requires understanding the business objective and the advantages and disadvantages of different predictor models. In doing so, it is critical to select an accurate yet interpretable model.

Despite their differences, models typically follow a similar modeling framework. First, the model splits the data into training, validation, and testing sets. Then, they can be used as classifiers to predict patient adoption of the drug.

Such models include:

• Logistics Regression Model: This model is ideal when the dependent variable is for dichotomous (binary). Logistics Regression describes data and explains the relationship between one dependent binary variable and one or more nominal, ordinal, interval, or ratio-level independent variables.³It is highly interpretable since the coefficients of the model indicate the most critical adoption drivers. However, this model is less accurate than others and suffers when higher-complexity data is input.
• Neural Networks (NN): Designed like the human brain, NNs consist of layers of many interconnected neurons. Each neuron, or “expert,” collates data from the layers beneath it, and passes information nonlinearly to the layers above it. In terms of accuracy, a typical NN outperforms other ML models. However, the model is sometimes difficult to interpret due to the several intervening layers. For this reason, many often refer to NNs as “black boxes.”
• Ensemble Model: Ensemble models combine the best predictions of several models for each patient to offset the inaccuracies of other models. Using “ensemble learning,” this model combines the strengths of several models to yield high accuracy. However, a mixture of different model types means potential difficulty in interpreting its results.

Analysts can also combine the best predictions from each model to gain higher-accuracy predictions using ensemble learning. Going forward, AI will continue to advance by increasing interpretability, and as a result, better predicting which drivers are responsible for decision-making

Use Case: Using DL to Predict Accurate Patient Adopters

Axtria, a big data and analytics company for the life sciences industry, estimated the chance of doctors prescribing patients specific drugs based on information from their medical claims. Using the achieved insights, Axtria helped commercial leaders better understand patient behavior and create high-targeted commercial strategies. Subsequently, the generated insights will increase the likelihood of drug prescription.⁴

Methodology

Typically, AI/ML advances use EHR data to extract detailed descriptions of the patients’ diagnosis, procedures, and drug administration. Instead, Axtria used claims data to identify patient characteristics, which EHR data might not provide.

Axtria’s steps were as follows:

1. Choose the best-suited predictor model.

• Create patient cohorts using unsupervised DL.
• Identify the objective behind insight derivation. In our case, the stakeholders wanted to understand the drivers of patients’ drug adoption better.
• Choose a prediction model. We selected the easily-interpretable Random Forest model to allow us to extract insights — a model similarly accurate to the Ensemble model.

2. Extract Insights from Claims Data

• Identify past treatment-based patient characteristic patterns from the claims data insights, including comorbidities, demographics, and their providers/HCPs.
• Derive patient characteristics from claims data, such as treatment patterns with anti-neoplastic agents, prognosis markets and tests, and disorders and other complications
• Utilize claims data insights to link patients to providers.

From the claims data, specific patient characteristics we identified were significant in predicting drug adoption, such as:

• Treatment patterns with anti-neoplastic agents.
• Prognosis markets and tests.
• Disorders and other complications.

Results

The approach delivered results with over 80% accuracy — a ~10% increase compared to previous methods, both in precision and recall. Specifically, the results determined that:

• Patients with a higher frequency of HCP-ordered tests were more likely to take the drug.
• The higher number of comorbidities a patient had, the less likely they were to take the drug.
• Patients’ regional differences and types of hospitals which they received treatment played a significant role in whether patients would adopt the drug.

Conclusion

Notably, patients’ regional differences and types of hospitals affected their adoption chances for several reasons. First, there is more substantial GPO influence in certain regions as compared to others. Second, certain academic hospitals have stronger research support and strive for clinical “championship.” Hence, they are more open to trying new drug treatments.

For sales leaders, the insights generated were critical to developing focused-messaging for HCPs, which help them make informed treatment decisions. Ultimately, these messages lead to an increase in prescriptions.

The Future of Deep Learning

Currently, AI technologies such as DL, extract insights much more efficiently than manual engineering. The latter is less useful, requiring expert domain knowledge which is sometimes unavailable. As compared to humans, DL closes these knowledge gaps and can anticipate patient drug adopters faster and more efficiently.

Providing useful insights about patient adoption of specific drugs is a critical business strategy-defining move for leadership. Moreover, these insights offer sales reps ammo during HCP interaction. For instance, the results provide them informed talking points about the drug, information on their effectiveness in treating specific disease states, and the like.

DL will continue to act as a catalyst in precision medicine. In the future, predicting adopters will require understanding specific patient types’ prescription patterns. Luckily, DL extracts useful insight from high-volume, high-variety data more efficiently and accurately than humans.

Going forward, understanding patient health will require a fusion of data sources, e.x. genomics, patient sensors (IoT), and detailed health records. Meaning, the next great task will include overcoming challenges from varied data standards, coverage, and other complications. However, AI tools will continuously advance to increase our understanding of such insights.

Interested in similar information? Read how Axtria used AI/ML to increase insights from unstructured data in EHRs.

Sources:

1. Wikipedia contributors. (2019, May 28). ‘Feature Engineering.’ In Wikipedia, The Free Encyclopedia. Retrieved 15:05, July 19, 2019, from https://en.wikipedia.org/w/index.php?title=Feature_engineering&oldid=899189316
2. Rajkomar, A., MD. (2018, May). ‘Deep Learning for Electronic Health Records.’ In Google Blog. Retrieved from https://ai.googleblog.com/2018/05/deep-learning-for-electronic-health.html
3. ‘What is Logistic Regression?’ (2019). In Statistics Solutions. Retrieved from https://www.statisticssolutions.com/what-is-logistic-regression/
4. Chickarmane, V., Bhamidipati, S. ‘Machine Learning Applications in Commercial Life Sciences.’ Presented by Axtria during the PMSA Winter Symposium 2019.