Improving Efficiency and Accuracy with AI in Healthcare

AI in healthcare is not new; in fact, it’s been trying to gain footing for years with two major challenges:

1. Doctors don’t trust a black box, and in order for AI to help deliver on its promise, trust (plus an explainable and white-box approach) is required to encourage buy-in from every technician, nurse, insurance provider, and clinician. 
2. Data in healthcare is sensitive, highly regulated, and thus challenging to work with in the bounds of privacy laws without the right tools and processes.

High-Value Use Cases

Fraud: Insurance fraud is a major issue that drives up healthcare 
costs for insurers, providers, taxpayers, and patients. Machine learning – specifically anomaly detection – presents a solid opportunity for the industry and regulators to fight back against fraudsters. By automatically combining hundreds of variables from different datasets, including patient/prescriber history, interaction graphs, prescription characteristics, and other contextual data, it became possible to target actual fraud cases.

Network analysis: Understanding relationships between physicians and facilities offers insights about patient behavior and medical costs that aren’t as easily gained through other analytical means, especially when combined with predictive modeling and machine learning.

Watch Video

Diagnosis support: Increasingly, medical providers are turning to clinical decision support systems (CDSS) to keep better tabs on patients, reduce errors, and lower costs. By advancing CDSS with AI, clinics and hospitals can diminish readmissions, as the system notifies a clinician about the patient’s electronic health record and flags information that may not have been considered, such as an allergy or a past negative reaction to a drug.

Patient care: Hospitals and other providers produce an immense amount of data. Until recently, this data has been largely unharnessed because it is unstructured and isolated from other relevant 
data points. Thanks to the development of innovative data-mining technology, healthcare providers can pursue revolutionary prevention and treatment strategies. Machine learning can help clinicians, medical providers, insurers, and other industry stakeholders understand and prevent dangerous conditions such as sepsis.

Operations: Virtual nursing assistants, automatic patient booking, and other AI-enabled tools have helped reduce healthcare costs. A major use case for machine learning in operations is predicting hospital bed or emergency room availability. Emergency room overcrowding is an expensive and dangerous problem that can be reduced by a system that predicts the number of incoming patients, and subsequently prioritizes discharging patients to free up space or redirects incoming cases to nearby hospitals.

Staffing optimization: Each hospital and clinic is dependent on its staff, whose salaries account for more than 50 percent of average hospital costs. In order to adapt staffing to daily needs, data platforms can drive better anticipation of patient demand (through patient forecasting), ensuring that there are enough health care providers to support a patient load.

Dataiku for Healthcare

Dataiku is the platform democratizing access to data and enabling healthcare organizations to build their own path to AI by:

• Making AI accessible to a wider population (including not just data scientists, but physicians, analysts, and business staff).
• Facilitating and accelerating the design of machine learning models to create AI-driven services.
• Providing a centralized, elastic, and governable environment that drives a responsible AI strategy.

Specifically, users and organizations in the healthcare space can leverage Dataiku for:

Automatic data processing and cleaning that can consolidate data from a variety of ordered (e.g., tabular) sources or sensor data. Pre-processing enables users to use “fuzzy matching” to clarify discrepancies in name and identity between different medical records.

Collaboration for technical, business, and medical users means that all stakeholders can visualize and understand the roots of data-driven insights, thus demystifying the “black box” and improving trust in AI-enabled systems.

Regulation-compliant access restricts which users can access what data, and pseudo-anonymization capabilities allows users without the necessary access permissions to engage with and learn from sensitive data without compromising patient privacy.

Subpopulation analysis equips users with tools to see how machine learning models respond to specific populations (e.g. age group, procedure type, prescription). This is particularly useful when exploring biases or anomalies within a system and can highlight areas for improvement in future updates.