Luis Miguel Chong Chong is a veteran business executive whose career has spanned decades across Mexico, Peru, and the United States, with leadership experience in information and communication technology, smart infrastructure, healthcare, and genomics. He currently serves as Managing Director of Disruptiva Works and as a partner in two food manufacturing and distribution companies, Grupo Ensi in Los Angeles and Megachef in Mexico City, while also sitting on the board as an external director of Grupo Viz-Sukarne, a meat industry holding company based in Culiacán. Throughout his career, Luis Miguel Chong Chong has focused on financing, developing, and operating major infrastructure projects such as highways and railroads, alongside a growing interest in renewable energy and broadband accessibility. Outside the boardroom, he is an accomplished judo competitor and former national junior champion. That same attention to rigor and evaluation informs how hospitals approach testing new AI tools before clinical use, the subject explored below.
Clinical AI tools analyze health information to support tasks such as reading medical images, estimating patient risk, and helping clinicians make decisions. Before using one in patient care, a hospital needs to know whether the tool is suitable for its intended patients, data, setting, and workflow. Testing therefore goes beyond asking whether the software works technically.
A tool may already have strong published results, but those results may not transfer directly to another hospital. Patient ages, disease prevalence, imaging equipment, image-processing methods, and local data practices can differ across health systems. These differences can change how well the same model performs in a new setting.
Hospitals first define the tool’s intended use. That means identifying the task it performs, the patients it applies to, the information it uses, who receives its output, and what clinical decision the output may affect. An imaging tool that flags possible abnormalities has a different purpose from a system that predicts which hospitalized patients are likely to see their condition deteriorate.
The next step is to test performance against an appropriate reference. Depending on the task, that reference may be a confirmed diagnosis, expert review, laboratory result, clinical follow-up, or another accepted standard. Testing data should also be kept separate from the data used to train the model so the evaluation measures performance on information the system has not already learned from.
One overall accuracy number may hide important differences in performance. Hospitals need to examine the types of errors the tool makes, including false positives and false negatives. A false positive flags a condition that is not actually present, while a false negative fails to flag a condition that is present, and those errors can have different clinical consequences.
Hospitals may also test whether performance differs across relevant patient groups. Evaluation can consider age, sex, race or ethnicity, medical condition, and other characteristics that matter for the intended population. A tool that performs well overall can still be less reliable for a subgroup that was poorly represented in the data used to develop or test it.
Technical performance is only part of the evaluation. Hospitals also test how the tool fits into clinical work, including when an alert appears, who sees it, and whether the result arrives in time to affect a decision. Studies of clinical decision support have shown that alerts can arrive after clinicians have already made a treatment decision.
Too many low-value alerts can also contribute to alert fatigue.
Human oversight is another part of safe use. Hospitals need clinicians and other responsible staff to understand what the tool does, what its limitations are, and how its recommendations should be reviewed. Organizations also need ways to record use, report problems, assign responsibility, and stop or disable a system if a serious safety concern appears.
A hospital can then decide whether the evidence is strong enough for routine use. The decision can include local performance, subgroup results, workflow testing, clinical-risk review, and governance requirements. If the evidence is not adequate, the hospital may need more testing, a narrower intended use, or a decision not to adopt the tool.
Pre-use testing also creates a baseline for what happens after deployment. Hospitals can compare later performance with the expected population, use conditions, and results established before launch. If patient demographics change, the model is updated, or performance begins to drift, that baseline gives the organization a basis for further review, restriction, or additional validation.
About Luis Miguel Chong Chong
Luis Miguel Chong Chong is a Mexico City-based business leader with decades of experience spanning information technology, smart infrastructure, healthcare, and food manufacturing across Mexico, Peru, and the United States. He serves as Managing Director of Disruptiva Works and holds partnership and board roles with companies including Grupo Ensi, Megachef, and Grupo Viz-Sukarne. His career has emphasized financing and operating major infrastructure projects, along with an interest in renewable energy and expanded broadband access. He is also a former national junior judo champion.
