Use Cases of Artificial Intelligence in Laboratories

As in many other sectors, artificial intelligence has become a topic of conversation in laboratories. Calibration and testing laboratories, however, are organisations whose work rests on the accuracy and traceability of results and which have little room for error. The subject of artificial intelligence in the laboratory should therefore be approached with caution rather than excitement. In this article we look at the possible use cases where artificial intelligence could be considered in laboratories, the risks that are specific to laboratory work and the orderly record keeping that all of this depends on. What follows is a general assessment. Each laboratory should consider the subject separately in the light of its own scope and rules.

The basic condition: orderly, digital and reliable records

Artificial intelligence tools work with data. If measurement results sit on paper forms, instrument details in separate spreadsheets and customer correspondence in personal mailboxes, there is no source for a tool to draw on. Talking about artificial intelligence before taking the step of laboratory digitalisation is like adding a floor to a building that has no foundation.

That foundation is usually provided by a LIMS (Laboratory Information Management System). A LIMS is the type of software that records, in one place, the whole process from accepting a job to delivering the result to the customer. Keeping records consistent, complete and clear about who created them and when is valuable to the laboratory itself, whether or not artificial intelligence is ever used. The order is clear: reliable records first, then the tools that may be built on top of them. Whenever LIMS and AI are mentioned together, it is worth remembering that the real effort belongs to getting the records in order.

Noticing unusual values in measurement data

Data analysis methods can be used to flag values that fall outside the usual pattern when compared with past measurements. An instrument showing an unexpected deviation from its previous calibrations, or a test result that clearly differs from similar samples, may be a situation worth reviewing. The cause may be a genuine change, but it may equally be a typing error, a wrong unit or the wrong instrument being selected. Done by hand, this kind of comparison takes time and can be skipped during busy periods.

The role of the tool here is to warn, not to decide. Whether a flagged value is wrong is decided by the staff who performed the measurement and know the method. A tool should not correct or delete a value on its own. It should also be kept in mind that not every warning will be right and that some real problems may pass without any warning at all.

Reviewing calibration intervals

How often an instrument is calibrated is usually determined by the manufacturer's recommendation, the conditions of use and the experience of the instrument's owner. If past calibration results are recorded in an orderly way, the behaviour of the instrument over time can be examined using that data. An instrument whose results have been stable for a long time and one that drifts noticeably each time may not be assessed in the same way.

Artificial intelligence and statistical methods can be considered as support for this review, for example by making a trend visible. The decision to change an interval, however, should be made by an authorised person together with its technical justification, and it should be recorded. Conclusions drawn from a small amount of historical data can be misleading, so caution is needed. Factors that do not show in the data, such as how heavily the instrument is used or whether it has been moved or repaired, should also be taken into account.

Extracting data from documents

The information that reaches a laboratory does not always arrive in an orderly form. Instrument lists sent by customers, request forms, external calibration certificates and scanned documents are often entered into the system by hand. Artificial intelligence tools could be considered for extracting details such as instrument name, serial number and measuring range from these documents and preparing a pre-filled draft.

This may reduce the burden of manual typing, but it also creates a new source of error. A tool may misread a digit or fill in information that is not in the document as if it were. Extracted data should therefore be treated as a draft to be checked rather than a final record, and approved only after comparison with the source document. It should also be understood that errors become more likely with documents that are hard to read, handwritten or laid out in an unusual way.

Drafting correspondence and reports

Large language models are strong at producing text. They can be used for the first version of texts such as an information e-mail to a customer, a description in a quotation, a draft procedure or internal training notes. For laboratories that correspond in a foreign language, translation and language correction may also be a useful area. Summarising a long correspondence history or a set of meeting notes can save time as well.

The line should be drawn where the measurement result and the technical evaluation begin. The results, uncertainty values and statements of conformity in a certificate or report are not texts for artificial intelligence to generate. They come from validated calculation and the evaluation of authorised staff. For draft texts too, the accuracy of the content is the responsibility of the person who signs or sends them.

Workload planning and answering customer questions

When past job records are in good order, it becomes possible to see which periods are busy, how long each type of job takes and which customers have instruments approaching their calibration date. This information can be used for forecasting in staff and equipment planning. Artificial intelligence is a forecasting tool here, and forecasts should be compared regularly with what actually happens. The plan is still made, and changed when necessary, by the laboratory manager.

On the customer side, the questions are mostly similar. People ask which stage their instrument is at, whether their document is ready or which scope of services is offered. An assistant answering these questions makes sense only if it relies solely on the laboratory's own up-to-date records. A tool that answers by guessing at information that is not on record can cause serious misunderstandings, especially on matters such as the scope of accreditation. Questions it cannot answer with certainty should be passed to staff.

Risks specific to laboratories

The product of a laboratory is a reliable result. Some risks that are acceptable for other businesses are therefore not acceptable in a laboratory. The following points should be observed when the use of artificial intelligence is being considered:

• The measurement result is not left to artificial intelligence: The result comes from a measurement made with a defined method and from validated calculation. A tool should neither produce nor alter a result.
• Traceability: It should be clear where a record came from and who created and approved it. What the tool suggested and what a person approved should be distinguishable.
• Validation: Like any tool used in the laboratory, an artificial intelligence tool should be tried out to confirm that it works as intended, and it should be understood that it may give different answers to the same input.
• Confidentiality: Customers' instrument and result information should not be entered into external tools whose terms are unknown.
• Responsibility: The person who signs the document and the laboratory are responsible. A suggestion from a tool does not remove that responsibility.

ISO/IEC 17025, the standard that defines the general requirements for the competence of laboratories, expects records to be traceable. Before bringing a new tool into their processes, accredited laboratories would do well to think about how it will be assessed in terms of their own quality system and accreditation expectations. Where there is doubt, the best course is to consult the current announcements and guidance of the relevant bodies.

Where could you start?

The healthiest route for laboratories is to start with tasks that are far from the measurement result and easy to check. Areas such as draft correspondence, summaries of internal documents or forecasts for planning do not touch the result, and mistakes in them are noticed easily. It is best to begin with a small trial, measure the outcome and put the rules of use in writing. Working with sample data rather than real customer data during the trial makes for a safer start in terms of confidentiality.

The step that comes before all of this is orderly record keeping. Regardless of artificial intelligence, keeping records in one place in a consistent and traceable way is what a laboratory needs today. To digitalise your laboratory processes, you can take a look at KALDATA Calibration Laboratory Management Software for calibration laboratories and Testing Laboratory Software for construction materials laboratories. To see the software up close, request a demo.