HHS released a playbook for trustworthy artificial intelligence in healthcare. Here are its ABCs, expanded upon and defined:
| (A)ccountability | Policies should outline governance and who is held responsible for all aspects of the AI solution (e.g., initiation, development, outputs, decommissioning) |
| (B)uilding an Algorithm | Humans can introduce bias when training an algorithm by selecting parameters that reinforce stereotypes or failing to engage diverse stakeholders (e.g., an AI project team that does not consult with members of affected communities may miss biased correlations in the algorithm’s decision-making process). |
| (C)omputer Vision | Intelligent algorithms that perform important visual perception tasks such as object recognition, scene categorization, integrative scene understanding, human motion recognition, material recognition, etc |
| (D)efining a Problem | Humans can introduce bias in the problem definition stage by selecting a target variable that is not a good proxy for the ideal target (e.g., an algorithm that predicts healthcare resource consumption may produce biased outcomes for individuals with less access to care). |
| (E)xplainability | All relevant individuals should understand how their data is being used and how AI systems make decisions; algorithms, attributes, and correlations should be open to inspection |
| (F)alse Positive/Negative Rate | The probability that a positive/negative result will be given when the true value is the opposite |
| (G)athering Data | Humans can introduce bias in the data preparation stage by using training data that is not representative of the target population (e.g., an algorithm trained on primarily White patient data may not perform well for BIPOC (Black, Indigenous, and People of Color) patients). |
| (H)uman Supervision | Human-in-the-loop: The AI solution provides recommendations, and a human reviews the output and makes a final decision Human-on-the-loop: A human monitors the AI solution and takes control when it encounters unexpected or undesirable events Human-out-of-the-loop: The AI solution has full control without the option of human override |
| (I)mpartiality | AI applications should include checks from internal and external stakeholders to help ensure equitable application across all participants |
| (J)ust-in-Time Learning | A concept where machine learning models adapt and update based on new data or experiences in real-time, allowing for continuous improvement. |
| (K)-Means Clustering | A popular unsupervised machine learning algorithm used for partitioning data into clusters based on similarity. |
| (L)arge Language Model | Large language models are advanced AI systems designed for natural language understanding and generation tasks. |
| (M)onitoring | Humans can introduce bias in the monitoring stage by using metrics that do not effectively evaluate fairness (e.g., evaluating overall performance of an algorithm without evaluating performance for affected subgroups may hide biased outcomes). |
| (N)atural Language Processing | Machines learn to understand natural language as spoken and written by humans |
| (O)verfitting | A modeling error that occurs when a function is too closely fit to a limited set of data points, resulting in poor predictive performance when applied to unseen data. |
| (P)rivacy | Individual, group, or entity privacy should be respected, and their data should not be used beyond its intended and stated use; data used has been approved by the data owner or steward |
| (Q)uality of Information | A multidimensional concept that encompasses critical relationships among multiple attributes such as timeliness, accuracy, and relevancy. Together, these attributes contribute to the validity of the information. |
| (R)obustness/Reliability | AI systems should have the ability to learn from humans and other systems and produce accurate and reliable outputs consistent with the original design |
| (S)afety/Security | AI systems should be protected from risks (including cyber) that may directly or indirectly cause physical and/or digital harm to any individual, group, or entity |
| (T)rustworthy AI (TAI) | Trustworthy AI refers to the design, development, acquisition, and use of AI in a manner that fosters public trust and confidence while protecting privacy, civil rights, civil liberties, and values, consistent with applicable laws |
| (U)se Cases | To determine whether a use case constitutes AI, consider whether the system… A. …performs tasks under varying and unpredictable circumstances without significant human oversight, or can learn from experience and improve performance when exposed to data sets? B. …uses computer software, physical hardware, or other technology to solve tasks that require human-like perception, thinking, planning, learning, communication, or physical action? C. …thinks or acts like a human, including the use of cognitive architecture or neural networks (e.g., developed to mimic the underlying mechanisms of the human mind)? D. …relies on a set of techniques, including machine learning, to approximate a cognitive task?E. …is designed to act rationally by utilizing intelligent software or an embodied robot to achieve goals using perception, planning, reasoning, learning, communicating, decision-making, and acting? |
| (V)alidation Data | A subset of the dataset used in machine learning that is separate from the training and test datasets. It’s used to tune the hyperparameters (i.e., architecture, not weights) of a model. |
| (W)hite-Box Testing | White-box testing in AI systems refers to a testing approach where the internal structure, design, and implementation details of the artificial intelligence model are known and considered during the testing process. |
| (X)AI (Explainable AI) | See (E). A field of AI focused on creating transparent models that provide clear and understandable explanations of their |
| (Y)ottabyte | A unit of digital information storage, representing one septillion bytes. In the context of AI, it relates to the massive amounts of data that can be processed and analyzed. |
| (Z)ero Knowledge Proofs | Proves that encrypted data are within given ranges without revealing any information about the data |
Health Bites 
Leave a comment