Decentralised molecular diagnostic workflow
A sequence of steps that brings molecular testing closer to the point of care, reducing the need to rely solely on a central laboratory. It includes sample preparation, amplification, data collection, interpretation and digital reporting.

AI/ML in diagnostics
A set of algorithms that can support result classification, detection of anomalous patterns, signal interpretation, quality control and operational monitoring. In a diagnostic device, AI must be validated and monitored throughout the product lifecycle.[11]

Ct curves
Fluorescence curves generated during real-time PCR. These are the raw data from which information such as target presence, threshold value and signal quality is derived.

Automated quality control
A set of software rules, internal checks and instrument monitoring that enable the flagging of questionable results, invalid runs or operational deviations before the report is issued.

Predictive maintenance
The use of operational data and analytical models to identify early signs of failure or instrument drift, reducing unplanned downtime and improving operational continuity.[9]

Human-in-the-loop
A model in which the system automates repetitive or analytical tasks, but final supervision and clinical responsibility remain with the healthcare professional.

Decentralised molecular diagnostics is entering a new phase. The first phase involved taking PCR out of the central laboratory; the second involved connecting the instruments; the third, today, consists of transforming the workflow into a data-driven system, where every curve, flag, log and metadata entry contributes to the overall quality of the process. The focus is no longer merely on performing a test close to the patient, but on managing a network of tests as a distributed clinical and industrial infrastructure.[7]

In this scenario, automation and AI are not mere add-ons. They are the layer that makes decentralisation scalable. Without management software, automated controls, standardised protocols and audit capabilities, a network of distributed instruments risks increasing variability and complexity. With a well-designed architecture, however, the same network can become faster, more controllable and more useful for clinical and organisational decisions.[1] The topic is particularly relevant for Helyx Industries S.p.A. because it touches on the heart of the new industrial structure: Hyris for distributed qPCR, Vytro for clinical and IVD applications, and Mytho for the NGS and bioinformatics dimension. The article focuses primarily on Hyris, because the decentralised workflow based on the Hyris System™ (bCUBE™, bAPP™, reagents) is the natural environment in which automation and AI can transform molecular data into repeatable and manageable processes. The reference to Vytro remains relevant as the same infrastructure supports clinical panels, IVD validations and hospital use.

The aim here is not to present AI as a generic promise. In diagnostics, AI is only useful if it solves very concrete problems: consistent curve reading, reduction of equivocal results, standardisation of quality control, device monitoring, interoperability, audit trails and risk management. For this reason, the correct question is not ‘how much AI is there in the workflow?’, but ‘which steps in the workflow become safer, faster or more controllable thanks to AI?’.

A decentralised molecular workflow is not simply a PCR test carried out on a smaller instrument. It is a process comprising physical, chemical, digital and organisational elements that must work together: sample, reagents, instrument, protocol, software, quality control, reporting and traceability. In the point-of-care models described in the literature, proximity to the patient is only useful if accompanied by clear procedures, interpretable results and the ability to integrate with the wider diagnostic system.[7]

Publications on bCUBE™ and Hyris rapid tests have demonstrated how a portable platform can be used in near-patient scenarios or outside a fully equipped laboratory, with results compared against reference methods. Studies on SARS-CoV-2 have evaluated the use of Hyris instruments and kits in rapid diagnostic settings, highlighting the potential role of portable platforms in increasing testing capacity and the accessibility of results. [4][5][6]

In the Hyris model, the value lies not only in the device. bCUBE™ represents the physical node; bAPP™ represents the software layer that enables data to be controlled, collected and organised; the reagents and protocols define the biological content of the test. This integration is the key point: if each component follows different logic, decentralisation increases complexity; if the stack is designed as a system, the complexity is absorbed by the platform and not passed on to the operator. Automation primarily acts on three steps. The first is the execution of the protocol, where the machine reduces variability linked to timing, temperatures and run management. The second is signal acquisition, where curves and logs are saved in a structured manner. The third is interpretation, where the software can apply consistent rules and flag doubtful or non-compliant results. In a distributed network, these three steps form the basis for comparability across different sites.

The cloud adds a further layer: it allows instruments and results to be viewed as part of a network, rather than as isolated units. This does not mean that every decision must be automated, but that every run can be tracked, monitored, archived and compared. Connected diagnostics has already proven its worth in public health programmes too, where diagnostic connectivity can transform laboratory results into actionable data for surveillance and decision-making.[8]

The most interesting aspect of AI in PCR is not the futuristic jargon, but its ability to interpret complex signals in a coherent manner. A Ct curve is not just a number: it is a time series, with trends, noise, an exponential phase, a plateau, any deviations and atypical patterns. Two curves with the same Ct value can have very different qualities. This is where data-driven analysis can add value compared to a static threshold alone. Recent literature on machine learning in point-of-care testing shows that AI models can be applied to diagnostic signals of various kinds – from optical tests to nucleic acid amplification techniques – to improve sensitivity, signal interpretation and operational accessibility. This promise, however, is accompanied by well-known challenges: model generalisability, training set quality, risk of bias, interpretability and validation in real-world settings.[1]

Within the qPCR domain, Pereira et al. described the development, validation and implementation of a multi-target qPCR method for HPV genotyping supported by software automation, data science and AI. The most interesting aspect is not only the use of the algorithm, but the fact that the automation has been validated in a clinical laboratory setting and applied to a high-throughput assay, where reducing manual intervention and interpretative variability has direct operational value.[3]

For Helyx Industries S.p.A., this issue is strategic because it links Hyris and Vytro. Hyris provides the distributed and connected platform logic; Vytro represents the clinical application layer, where IVD kits and panels require standardisation, robustness and clear reporting. From this perspective, AI is not a mere embellishment: it is a means of safeguarding data quality when the test leaves the central laboratory and must remain interpretable, comparable and auditable.

In diagnostics, quality control is not a mere formality. It is the very condition that makes the result usable. In a centralised laboratory, many deviations are detected by experienced staff, established routines and manual review.

In a decentralised system, this safeguard must be built into the workflow: internal checks, software rules, alerts, audit trails and remote supervision must function as a safety net. An automated quality control module can operate on multiple levels: verifying correct protocol execution, curve analysis, checking for the presence or absence of expected signals, identifying patterns inconsistent with true amplification, batch and operator tracking, and flagging borderline results. These elements do not eliminate human validation, but they reduce the number of decisions left solely to individual experience. This is a decisive step towards scalability. If each peripheral site interprets the result according to local practices, the network becomes fragmented.

If, on the other hand, quality rules are codified and distributed via a common platform, the central laboratory can maintain scientific oversight whilst leaving execution close to the point of care. It is the difference between weak decentralisation and governed decentralisation. The REASSURED framework, proposed for decentralised diagnostics, encompasses not only accuracy, speed and accessibility, but also connectivity, ease of use and robustness. These criteria are relevant because they remind us that point-of-care testing must not only be small and fast: it must also be safe, interpretable, usable and integrable into real-world healthcare systems. [10]

In the case of the Hyris System™, the combination of bCUBE™, bAPP™ and standardised protocols meets this requirement precisely: to make the test distributed without making it isolated. Quality control thus becomes a platform function, not merely the responsibility of the individual operator.

The benefits of AI and automation must be translated into KPIs; otherwise, they remain mere technological rhetoric. In the molecular diagnostics workflow, the most relevant metrics are TAT, failure rate, repeat tests, uptime, validation time, manual workload and data quality. Not all of these are always published as comparative metrics, but they all impact the operational sustainability of the decentralised model.

TAT is the most visible KPI. In point-of-care models, reducing transport, batching and waiting times at the central laboratory can shorten the time between sample collection and decision. Automation can contribute on a second level: by reducing manual steps, standardising quality controls and flagging doubtful or non-compliant results within predefined rules. This does not mean bringing forward the report, but making the path from raw data to validation more efficient and traceable.[1] [2]

The failure rate is less visible but equally important. Every failed test consumes reagents, time, instrument capacity and trust. Automation can reduce certain operational causes of failure, whilst software quality control can detect anomalies before they result in unusable reports. The experience described by Pereira et al. in the context of multi-target qPCR demonstrates how data science and automation can improve the management of complex datasets and reduce manual intervention in the analysis. [3]

Uptime is the KPI of greatest interest to those managing a network of instruments. A single machine out of action can be a problem; a distributed fleet without monitoring can become an organisational risk. Predictive maintenance applied to medical equipment is based precisely on the continuous analysis of operational signals to identify patterns of failure or degradation before service is interrupted.[9]

ROI should not be viewed simply as a comparison between the cost of decentralised testing and the cost of centralised testing. The economic value stems from the process: shorter waiting times, fewer repetitions, reduced logistical dependence, better use of staff, reduced downtime and greater data availability. In healthcare, the cost of a delayed result can be higher than the cost of the reagent. For this reason, automated diagnostics should be assessed as operational infrastructure, not merely as a consumable.

A traditional laboratory often only discovers a problem once a failure has already occurred: an optical channel is not performing, a temperature is not reaching the setpoint, or software is not recording data correctly. In a distributed model, this risk increases because the instruments are not all physically close to the technical team. Continuous monitoring serves precisely to transform every device into an observable node. Predictive maintenance uses operational data – temperatures, ramp times, optical signals, error logs, utilisation cycles, frequency of anomalies – to estimate when a component or process is deviating from an expected range. Shamayleh et al. describe an IoT and machine learning approach to predictive maintenance of medical equipment, based on the use of real-time data to anticipate failure modes and support maintenance decisions. [9]

Applied to a distributed qPCR network, this logic has clear value. If a device at a peripheral site shows progressive drift, the system can flag it before it produces a series of questionable runs. If a group of instruments shows recurring patterns, the quality team can investigate the batch, protocol, environmental conditions or operational behaviour. Technical data thus becomes a tool for prevention, not just for fault diagnosis.

Here too, AI does not replace procedures. It reinforces them. Scheduled maintenance remains necessary; AI allows it to be more targeted. The difference lies in moving from a blind schedule to a model where the schedule is supplemented by evidence of actual usage. For a laboratory or healthcare network, this means reducing the risk of unexpected downtime and making the planning of diagnostic capacity more reliable.

Automation and AI do not eliminate the laboratory. They transform it. The central laboratory remains the bastion of expertise, quality and scientific accountability; what changes is the distribution of operational activities. Some of the execution moves closer to the point of care, whilst validation, supervision, quality control and aggregate analysis remain centrally managed.

This shift alters the skills required. The technician is no longer merely a manual performer of repetitive tasks; they become a process manager, workflow supervisor and data interpreter. Familiarity with dashboards, flags, audit trails, alerts, algorithmic outputs and escalation procedures is required. At the same time, personnel from the laboratory, IT, quality and data science must work together. Diagnostic connectivity adds a further layer. Mujuni et al. demonstrate how connected digital platforms can go beyond simply transmitting results, generating actionable data in real time for infectious diseases and public health. The message is also applicable to the industrial context: a connected diagnostic result is more useful than an isolated result, because it can inform monitoring, quality control and organisational decisions.[8]

The risk is believing that technology alone is sufficient. It is not enough. An automated network without governance generates new risks: uncontrolled access, non-standardised data, ignored alerts, unclear responsibilities. For this reason, an AI-enabled molecular diagnostics project must include procedures, data ownership, training, deviation management and business continuity plans. Within this framework, Helyx Industries S.p.A.’s One Group – Three Divisions model helps to correctly interpret the positioning. Hyris oversees the distributed platform; Vytro can translate that platform into high-value IVD clinical applications and panels; Mytho oversees a different domain, that of NGS and bioinformatics, where the data logic is even more complex but should not be confused with PCR. Industrial coherence stems precisely from distinguishing between these areas of expertise and then enabling them to interact.

The shift from supporting software to a component that influences the diagnostic outcome changes everything. When an algorithm interprets a signal, flags an anomaly or contributes to validation, it cannot be treated as a mere computational function. It must be designed, validated, documented and monitored as part of the diagnostic system. The IMDRF’s Good Machine Learning Practice establishes principles for the development of safe and effective AI/ML medical devices throughout their entire lifecycle. Key points include data quality, appropriate training and test sets, risk management, independent validation, transparency and post-market monitoring.[11]

In the United States, the FDA has developed a dedicated pathway for Predetermined Change Control Plans for AI-enabled software functions, precisely because models can evolve over time. The logic is clear: iterative innovation is possible, but it must be anticipated, controlled and accompanied by a methodology for development, validation and impact assessment.[12]

In Europe, the AI Act introduces a horizontal framework for artificial intelligence systems and intersects with the MDR and IVDR when AI is incorporated into medical devices or software. Document MDCG/AIB 2025-6 clarifies precisely the interaction between medical device regulations, the IVDR and the AI Act, including in vitro diagnostic devices with AI components.[13] [14]

Transparency is another critical issue. The FDA/Health Canada/MHRA guiding principles for ML-enabled devices emphasise the need to clearly communicate intended use, performance, limitations, data used and update logic. In practice, the user must know not only that the system uses AI, but also what the AI does, when it is reliable and when escalation is required. [15]

Finally, there is cybersecurity. A connected diagnostic network expands the attack surface: tools, the cloud, data, interfaces and APIs all become part of the clinical risk. The FDA highlights the need for threat modelling, security architecture and documentation throughout the device’s lifecycle. For a distributed diagnostic platform, cybersecurity and diagnostic quality are no longer separate issues.[16]

Q: When it comes to AI in molecular diagnostics, where do you see the greatest risk of misunderstanding?

Lorenzo Colombo: The most common misunderstanding is to think that AI is a magic button. It isn’t. In molecular diagnostics, value comes from good data, robust controls and well-designed processes. AI can detect patterns that would be difficult for an operator to spot, but if the workflow is fragile, the algorithm won’t save it. That’s why we always think in terms of the system as a whole: the instrument, reagents, software, quality control and supervision must all work together.

Q: At what point does automation really change the way a laboratory works?

Lorenzo Colombo: It changes the relationship with reproducibility. In a traditional laboratory, much of the quality depends on people’s experience, which remains fundamental. But when you roll out the test to multiple sites, with different operators and conditions, you need to incorporate some of that quality into the process. That is what automation is for: reducing ambiguous steps, making decisions traceable and allowing the central laboratory to maintain scientific oversight even when the work is carried out elsewhere.

Q: Will AI be able to replace human validation?

Lorenzo Colombo: No, and that is not the aim. Human validation remains the point of responsibility. AI can help to better detect weak or anomalous signals and interpret them more consistently. It can flag an anomalous curve, a weak pattern, or a possible instrument drift. But the final decision, especially when it has clinical implications, must remain within a clear professional framework. I see AI as a co-pilot: it enhances capability and attention, it does not replace the pilot.

Q: Why is the cloud so important in a decentralised workflow?

Lorenzo Colombo: Because without the cloud, the risk is having many small, isolated laboratories. With the cloud, however, every node operates within a common framework: the same protocols, the same controls, the same interpretation criteria, and data available for auditing and improvement. The cloud is not just storage. It is the way in which you transform a constellation of tools into a governed diagnostic network.

Q: How does this fit into the new three-division structure at Helyx?

Lorenzo Colombo: Hyris is the heart of the distributed platform: hardware, software, reagents and workflow. Vytro brings this capability to clinical applications, panels and IVD kits. Mytho operates on a different level—that of NGS and bioinformatics—but shares the same data culture. The important thing is not to confuse the boundaries: each division has a clear mission. The group’s value lies in the fact that these missions can interact without overlapping.

Q: What is the next credible leap forward?

Lorenzo Colombo: Rather than a single leap, I see a convergence. More stable instruments, more robust reagents, smarter software, more explainable AI, and more automated quality control. The aim is to make molecular diagnostics more predictable and more scalable. It is not enough to be fast: we must be reliable in the same way, in different places, with different people and different workloads.

Automation and AI are transforming decentralised molecular diagnostics from a collection of portable devices into a governed digital infrastructure. The difference is substantial: a device performs a test; a platform controls the process, records the data, applies quality control rules and enables supervision at scale. The available evidence points to a clear trajectory. Machine learning can support signal interpretation at the point-of-care; data science can automate multi-target qPCR analysis; connectivity can transform local results into actionable data; predictive maintenance can enhance operational continuity and reliability.[1][3][8][9]

The limitation is equally clear: in diagnostics, AI cannot be treated as a purely commercial element. It must be validated, monitored, explainable and integrated into a quality system. The GMLP guidelines, the FDA pathways for predetermined changes, the AI Act and the MDCG/AIB guidance confirm that the future of medical AI will be measured as much on performance as on governance.[11][12][13][14]

For Helyx Industries S.p.A., this context involves Hyris, distributed qPCR, the cloud, data quality and the platform. But it is also a useful opportunity to ‘read’ the entire industrial structure. Vytro benefits from the same infrastructure when the workflow becomes clinical and IVD; Mytho demonstrates that the group is building expertise even at the most advanced level of genomic data. One Group – Three Divisions means precisely this: distinct scopes, but a common logic of industrialised molecular biology. The new frontier is designing diagnostic workflows in which every stage – from the Ct curve to the report, from quality control to maintenance, from the individual device to the network – becomes more traceable, more standardised and more reliable. In this sense, automation and AI do not replace the laboratory: they extend it.