Clinical Decision Support System can be the biggest challenges in AI and Healthcare

In the Healthcare industry, as the magnitude of data piles up, so does the necessity of Clinical Decision Support (CDS) Systems. With the rising trend of value-based healthcare, clinicians must have tools at their disposal that can help them sift through mountains of records that can sometimes date back to decades.

According to a report, a third of deaths in the US are caused due to medical errors and inaccuracies. This justifies the need for deeper analysis and reporting of accessible data on the patient, with members of the healthcare community. Although one cannot eliminate these errors from the domain, having a system that minimizes these acts as a safety net. Hence CDS is a must for a paradigm shift in this industry.

Since the 1980s when it was first used, it has quite evolved much. Now it is administered via Electronic Health Records (EHR) and computerized provider order entry (CPOE) systems. This has been made possible by increasing the global adoption of HER with advanced capabilities, e.g. FitBit.  These systems comprise of 3 distinct parts: Base (pre-defined algorithms), Inference Engines (output is generated based on patient's clinical data) and Communicator (where interaction is made with user/patient). The only challenge one can face here is poor implementation as it will lead to alarm fatigue and clinician burnout. While benefits encompass safety of patients, increased adherence to guidelines and care pathways, prevention of Drug-Drug Interactions (DDI) errors, cost containment by suggesting cheaper alternatives, support for clinical and diagnostic coding, imaging, ordering of procedures and tests, and patient triage, qualitative documentation, etc., certain barriers need to be addressed by stakeholders.

Disruptive Workflows: Constant alerts and documentation, can unnecessarily create a fragmented environment for the practitioners. This can cause to put more cognitive efforts, and time to complete those tasks.  Even if there is time for face-to-face interaction with patients, chances remain that it can be abruptly cut off for the same reason.

Data Collection: Due to security and privacy reasons, patients may not permit the accessibility of their health or personal data required for diagnostics. This can be a hurdle when trying to trace data or data that exists in separate systems.

Alert Fatigue: Often, people in the medical industry are bothered by incorrect or unnecessary alerts. This includes false alarms about patients, too much information overload, etc. Hence, it is essential to build a model of specialty-specific, relevant standards to prevent the poor signal to noise ration caused by CDS. In addition to that, the system should be able to distinguish between scenarios that need immediate attention and ones don't require professional help. Else it can lead to alert fatigue when physicians pay heed to every alert.

User Skill: Because of the very promising nature, it is likely possible for the user to have too much reliance on the system and exploit it more than the expected limit, not cross-check for the correctness of the data, etc.

Regulation: In case of malpractice cases involving medical Machine Language applications arise, the legal system needs to provide clear guidance on what entity holds liability, however, regulations in this arena still lag. Furthermore, ML in healthcare poses a unique challenge to regulatory agencies because the models can quickly evolve as more data and user feedback are collected, and it is not clear how the updates should be evaluated (FDA, 2019)

Bias vs. Fairness: Even a machine or system cannot be free from discrimination. Pieces of evidence prove that CDS can refuse to adapt or accept a new environment (for test recommendation, treatments, etc.) that differs from its original algorithm. This can be based on country, region, or sex or age.

System and Content Maintenance:  In the CDS life-cycle, maintenance is the most neglected part despite its huge importance. This includes technical maintenance of systems, applications, and databases that power the CDS. Even it is mandatory that the knowledge base is kept in pace with the ever-changing medical world and clinical guidelines.

Lack of Interoperability: Commonly CDS exists as standalone systems or in systems that aren't able to communicate with other systems, although cloud storage can help solve this issue.

Verification and validation: For successful adaption and running of clinical rules of CDS used, it is important to test all functions of the software products. Tendering, choosing or implementing a new CDS requires a comprehensive user requirement specification (URS) or user requirement documentation (URD).

CDS have been shown to enhance healthcare providers in a variety of decisions and patient care tasks, and today they actively and ubiquitously support the delivery of quality care. If it manages to locate the problem and understand qualitative and meaningful data, its achievements can belong terms. Hence, it is important to make it a joint and closely integrated effort to handle the challenges specified earlier. On the algorithms themselves, a bulk of research should be dedicated to dealing with longitudinal data, how to best describe a patient, and how to relate the learning with pathophysiology, i.e., how can we marry previous clinical knowledge with the algorithmic conclusions. Data integration and what is the best approach for dealing with incomplete data and outliers should be also surveyed. One should also ensure that practical usage of this technology does not burn a hole in the pockets of common mass.