Table 4.

Challenges with machine learning models in healthcare.

Aspects of Machine Learning-models and the healthcare environment	Gaps or Challenges
Context	- Need to thoroughly understand the clinical data being used for model development (13, 36) [C1] - Need models with impactful clinical utility (13) [C2] - Need models that fit within the environment they are intended for (13, 37–41) [C3]
Data	- Need access and availability to well labeled, high quality, large datasets (13, 14, 39) [C4] - Need consistency in data collection techniques (13) [C5] - Need to acknowledge and minimize inaccurate or incomplete data (13, 41) [C6] - Need to ensure that model training/test data is representative of what the model will experience during operation; consider pre-processing of data and its effect [C7] - Need to identify, remove, and account for biased data (13, 14, 40, 41) [C8] - Need to account for data shifts and their effect on model performance [C9]
Model validation and performance	- Need to conduct and develop clinical validation studies (11, 13, 14, 37) [C10] - Need to conduct clinical impact/outcome studies as Machine Learning metrics (accuracy, precision, etc.) often do not map directly to clinical performance indicators (14, 37) [C11] - Need model transparency (11, 39, 41) [C12]
Ethics and Regulation	- Need regulation and safe use guidelines (14, 39, 42) [C13] - Need privacy and cybersecurity regulations (39–41, 43) [C14] - Need to screen for algorithmic biases (11) [C15]
Financial issues	- Need adequate resources (hardware, expertise, software, etc. all in high demand, limited, and expensive) to develop and integrate models (39) [C16]
Knowledge gap	- Need users to have sufficient knowledge to interpret model output or compare different models (11, 39, 41, 44) [C17]

ML, Machine Learning.