Deterministic automation
Software that applies predefined rules, for example by invalidating a run when a required control fails to amplify.

Machine learning on the curve
A model that recognises patterns in the fluorescent signal; performance must be demonstrated for the intended assay, platform and population [2,4].

Intended purpose
The stated function of the software, with its inputs, outputs, users, population, context of use and the decision it influences clearly defined [4,11,12].

External validation
Verification on independent data and under representative conditions, including instruments, lots, sites, matrices and subgroups that the model will encounter in its intended use [4,6–8].

Change control
Governance of versions, modifications and updates, including impact assessment, acceptance criteria, monitoring and the ability to fall back [4,9,10].

In molecular diagnostics, ‘artificial intelligence’ can refer to very different objects: a curve classifier, a multi-site quality-control system, an interface that prioritises cases for review, or a dashboard that aggregates already validated results. The distinction affects the type of error that may occur, the evidence required and the responsibility associated with the output.

This Insight distinguishes three levels—signal interpretation, process surveillance and epidemiological interpretation—and examines them through applications in infectious diseases. More than establishing whether AI is ‘better’ than conventional interpretation, the task is to understand when an algorithmic function becomes assessable, verifiable and useful without concealing uncertainty.

The ‘AI-driven’ label is applied to different phenomena: a thermocycler connected to the cloud, an automatically calculated threshold, software that verifies controls, a machine-learning model that classifies a curve, and a dashboard that displays the geographical distribution of positive results. These are potentially useful functions, each requiring different evidence.

A rule that invalidates a run when the internal control fails to amplify is deterministic automation: the criterion is predefined and the software applies it. A model that learns from thousands of curves to distinguish plausible signals, artefacts and uncertain cases belongs to machine learning. A map that aggregates already validated results does not interpret the test; instead, it supports epidemiological interpretation. Treating all three levels as ‘AI’ increases the impact of the claim and reduces the ability to assess them.

The useful question concerns the input received by the algorithm, the output produced, the decision influenced and the expected behaviour in the event of an error. This is the step that brings the discussion back from marketing to diagnostics.

Real-time PCR generates a time series of fluorescence measurements. The Cq—the cycle at which the signal crosses a threshold—condenses that dynamic into an operational value, but it does not on its own describe shape, slope, background noise, plateau, control behaviour or interactions between channels. MIQE 2.0, a guideline for reporting qPCR experiments, requires transparency regarding pre-analytics, assay design, efficiency, limits of detection and quantification, controls and raw-data availability [1]. It is not a standalone standard for clinical validation; it does, however, make what the workflow measured and documented verifiable, which is also a necessary condition for assessing an algorithm.

A frequently cited example is qPCRdeepNet, a model developed to interpret fluorescence readings from a SARS-CoV-2 assay without relying exclusively on Cq. In the dataset studied, analysis of the complete curve showed potential as a quality-assurance tool and for identifying anomalous results [2]. The finding confirms that the signal contains usable patterns, but it remains limited to that assay and those conditions: transferability to another assay or thermocycler, a different matrix, or a pathogen with different dynamics must be demonstrated.

This is where a decisive boundary lies: in diagnostics, strong retrospective performance on internal data begins validation; it does not complete it.

The most direct use case concerns curve classification. A model can assist in distinguishing positive, negative, invalid or uncertain results; recognise patterns incompatible with the expected amplification; highlight late or irregular curves; and recommend priority review. In multiplex panels, it can also consider relationships between targets and controls instead of assessing each channel in isolation [2,4].

The potential operational benefit is consistency: in high-volume laboratories and multi-site networks, a well-validated system can apply the same criterion and make exceptions more visible. The correct verb, however, remains ‘assist’. The meaning of the result also depends on what the curve does not show: contamination, sample quality, inhibition, stage of disease, therapy, prevalence and patient history [1,3,4].

A study of highly automated NAAT platforms for SARS-CoV-2 estimated a very low proportion of false positives in the cohort analysed; the authors nevertheless limited the generalisability of the estimate to other settings [3]. When prevalence is low, even rare errors can affect the probability that a positive result is truly positive. Software can recognise an unusual curve; the missing clinical information remains unresolved.

A robust workflow retains a category of uncertainty. Forcing every case into ‘positive’ or ‘negative’ may appear efficient, but it removes from review precisely those samples that require repetition, an orthogonal method or specialist assessment [4].

AI can observe series of runs and different sites to recognise changes that an individual test does not reveal: an increase in invalid results, a shift in the Cq distribution, rising noise in one channel, or differences between lots, instruments or operators. At this level, the unit of analysis becomes the process rather than the individual patient [4].

Connectivity and standardisation are crucial. Comparing data from a network requires consistent metadata: software version, reagent lot, instrument, assay, matrix, extraction procedure, controls, environmental conditions and technical interventions. Without this structure, the model may detect a variation without knowing whether it represents an analytical problem, an epidemiological change or a simple workflow modification [1,4].

The distinction between biological and technical drift is particularly important in infectious diseases. An increase in positivity may indicate an outbreak, contamination or a new criterion for referring samples. A mutation in the target may reduce assay performance; a reagent change may alter curve shape. An alert is useful when it leads to a reproducible investigation, not when it merely produces more notifications [1,4].

International good machine learning practice principles therefore indicate datasets representative of the intended use, adequate separation of training and test sets, assessment of human–AI interaction, clear information for users and lifecycle monitoring [4]. The algorithm operates within the quality system, without replacing its responsibilities and controls.

A network of distributed tests can generate a useful geotemporal flow. By aggregating results with date, location, target and reference population, analytical tools can recognise unexpected changes, clusters or geographical differences. The potential benefit concerns the timeliness of collective interpretation, not the accuracy of the individual PCR test [5].

Epidemiological quality, however, depends on the denominator. Ten positive results can mean different things if they come from twenty tests in symptomatic people, one thousand screening tests, or a newly activated site. If access, sampling criteria or capacity change, the observed series changes too. Reviews of AI early-warning systems highlight problems involving data volume, granularity, availability, bias and adaptability [5]. Without context, a dashboard turns testing into a map, but not necessarily into surveillance.

Minimum clinical data, access rules, interoperability, information protection and a clear separation between the individual result and aggregate use are therefore required. Data speed becomes valuable only if authorities or healthcare organisations have thresholds, expertise and procedures for acting [5].

In respiratory infections, automated curve analysis can help manage high volumes and multi-target panels. The qPCRdeepNet study demonstrates feasibility in the dataset and assays assessed, without establishing general transferability [2]. External validation across seasons, variants, sites and different prevalences therefore matters more for adoption than an isolated accuracy percentage.

In sexually transmitted infections, the use case should be defined by taking account of co-infections, low-concentration targets and different clinical consequences. An algorithm can make the application of criteria and the review of controls more consistent; it cannot turn a detected target into active disease without considering the matrix, natural history, diagnostic window and applicable guidelines. These are conditions to be verified in the intended purpose, not a benefit that has already been demonstrated [1,4].

For potential use in sepsis or hospital-acquired infections, the assessment must extend beyond curve interpretation to include collection, volume, preparation, concentration of the agent, contamination and the distinction between colonisation and infection. Faster classification alone does not demonstrate an improvement in sample-to-answer time or outcomes. AI can be assessed for triage and process control. Any clinical benefit must be demonstrated across the complete workflow and against endpoints defined in advance [1,4,6–8].

A practical rule emerges: the greatest value does not always coincide with the most spectacular algorithmic function. Flagging an anomalous run early or avoiding an unnecessary repeat may be more useful than generating a complex interpretation.

A credible assessment begins with the intended use: population, sample, assay, instrument, operator and decision. Data and metrics follow from these variables. Sensitivity and specificity may not be sufficient; calibration, subgroup performance, behaviour in out-of-distribution cases, the human override rate, concordant and discordant errors, time saved and the consequences of errors must also be considered [4,6–8].

TRIPOD+AI, CONSORT-AI and SPIRIT-AI are reporting guidelines, not validation standards. They make models, protocols and trials more transparent by requiring descriptions of the data, the intervention, human–AI interaction and errors [6–8]. The principle is also useful in diagnostics: the evidence must allow the model to be assessed as a component of a pathway, without attributing to these checklists a regulatory or validation function they do not have.

External validation must include conditions that the model will actually encounter: different instruments, new lots, operators, matrices, co-infections, prevalences and varying sample quality. After release, it is necessary to observe whether the data distribution changes. A ‘locked’ model can lose performance; an updatable model introduces the opposite problem, because each change can alter the risk profile [4,9,10].

In 2025, the FDA proposed a total product lifecycle approach for AI-enabled software functions in a draft guidance; in the same year, it published final guidance on predetermined change control plans for planned modifications [9,10]. The two documents have different statuses, but they converge on one operational point: learning and updating require transparent versions, limits, monitoring and change control.

Applied to PCR, AI can add value to data that are currently compressed, discarded or observed too late. It can make interpretation more consistent, identify weak signals in the process and connect a network. Its value increases when the boundary between what the algorithm knows and what requires context remains visible [2,4,5].

For a laboratory or healthcare network, asking ‘do you have AI?’ is of little use. An assessment should instead establish which errors the function is intended to reduce, on which data it has been validated, how it manages uncertainty, what it shows the user, how it is monitored and which decision it is intended to support. The answers distinguish a digital function from a diagnostic capability [4,9,10,12].

AI-driven PCR should be understood as an architecture supporting a molecular process rather than as a standalone category of test. Its maturity is measured by how precisely it separates signal, process and population; by data quality; by its ability to state uncertainty; by external validation; and by the discipline with which updates are managed [4,9,10].

The most credible future avoids handing the diagnosis to an invisible algorithm. It focuses on systems in which automation and machine learning make the workflow more observable, direct attention to relevant cases and produce usable data. Intelligence emerges from the whole architecture, including the rules that establish when to trust, when to verify and when to stop.