Predictive Models Development to Support Fast-track Clinical Decisions for Healthcare

NIX built a robust data analytics solution with real-time predictive models, empowering healthcare providers to enhance patient outcomes, improve diagnosis accuracy, and reduce costs.

Business Domain

Healthcare
Service

Data Engineering
Technologies

Python, Tableau, Jupyter

Business Overview

The company is a large U.S. provider of multiple software solutions for the healthcare sector that help optimize costs, reduce risk, support compliance processes, enhance customer engagement, and create new revenue streams.

One of the client’s products is a web platform that helps hospitals and insurance companies analyze and predict the likelihood of various events by running sophisticated predictive models based on historical data, such as:

Cost efficiency of treatments compared within the industry
Risk of mortality based on a patient’s condition, diagnosis, and treatment plan
Identification of common trends in discharges, treatment regimens, and other factors among diverse patient populations
Complication predictions based on patient condition and similar historical data of other patients
Prediction of how long specific patients will stay at the hospital to forecast hospital costs

The client was looking to extend the offering with new analytics, expand the coverage to include additional hospitals in the U.S., and provide the healthcare industry with sophisticated analytics to enhance patient outcomes. To make this possible, the company approached NIX—a proven data science development and consulting company—to enhance the existing predictive models and ensure their smooth performance in the long run.

Challenge

The existing predictive models were developed many years ago and needed to be updated annually due to constant changing of rules in the healthcare industry such as treatment plans, diagnosis definitions, new procedures, etc. Failure to comply with regulations in the healthcare industry can lead to various negative consequences such as financial setbacks, breaches in security, revocation of licenses, disruptions in business operations, subpar patient care, loss of trust, and damage to reputation.

Therefore, the NIX team needed to train models and keep them up to date, ensuring quality monitoring and re-training with newer historical data.

Solution

The client provided the NIX team with a modeling Informix dataset to train predictive models—one of the largest all-payer inpatient databases. This dataset covers several years’ worth of data and includes over 60% of all discharges from U.S. inpatient hospitals. It’s a compilation of information from both public and proprietary state sources, as well as from individual and group hospital contracts.

The NIX process included the following steps:

01
Data Preparation

Transforming and categorizing the data by mapping the input diagnosis, procedure codes, discharges, etc.
02
Data Cleaning

Identifying and correcting errors or inconsistencies in the data such as missing or incorrect values and duplicate records using SAS—a programming language used for statistical analysis. This excludes hospitals that don’t match our models (federal-owned facilities, non-US states facilities, etc) as well as types of discharges (against medical advice, without official leave, etc).
03
Feature Selection

We identified the most pertinent features or variables that are expected to impact the predictive model. In doing so, we established discharge attributes that influence our events, such as the relationship between neonatal mortality and birth weight, or the presence of non-standard diagnoses. This helped minimize the data’s dimensionality, improve the model’s precision, and mitigate overfitting.
04
Model Training

By utilizing Python (Numpy, sklearn, SciPy, pandas), we trained the chosen models to detect patterns, correlations, and relationships in the data that can aid in making predictions.

We utilized different types of feature selection processes to identify a set of predictors, and subsequently ran logistic regression to gauge the influence of these predictors on our events. Finally, we selected parameters and a set of features to optimize the performance of our predictive model.
05
Model Evaluation

The trained model’s performance was assessed by utilizing Tableau for data visualization, SAS business intelligence tool, and Python to measure model’s performance. This evaluation process facilitated the identification of areas for improvement and the determination of the model’s overall effectiveness.
06
Model Deployment

We deployed trained models to a production environment where it can be used to make predictions on the сlient’ss data.

Outcome

The client received a robust data analytics solution with up-to-date predictive models that can be integrated with other healthcare systems and tools to provide real-time insights and recommendations to healthcare providers.

Healthcare providers have fast access to evidence-based clinical decision support systems, leading to:

Improved patient outcomes: By identifying early patients who are at high risk of developing diabetes, hospital readmission, or post-surgical complications, hospitals can take proactive steps to prevent or manage these conditions.
Improved diagnosis accuracy: Predictive models help healthcare providers more accurately diagnose patients by analyzing large amounts of patient data and identifying patterns that may not be immediately apparent to a human.
Reduced costs: Predictive models help lower healthcare expenses over time by enabling earlier detection and more individualized treatment.