Clinical decision support Medication monitoring Protocol in the workflow

Custom clinical decision support software

Decision support software alerts a healthcare professional at the moment it matters: an interaction between two drugs, the protocol step that fits this clinical picture, a lab value that falls outside the agreed range, an incomplete record. The starting point is fixed and does not change anywhere on this page: the healthcare professional makes the diagnosis and takes the treatment decision, while the software brings information to their attention at the right moment. Appfront builds that alerting layer to your requirements, alongside or on top of the system where the work already happens.

What clinical decision support is and is not

Clinical decision support is the umbrella term for software that brings together recorded knowledge and data from the record and turns it into something visible at a specific moment. In practice, the applications fall into four groups. Medication monitoring checks a prescription for interactions, duplicate medication, dosage, contraindications and hypersensitivities. Protocols and guidelines show the next step at the moment that step is relevant, rather than in a document someone consults afterwards. Alerting looks at measured values and combinations of values and signals when they fall outside an agreed range or change too quickly. Record checks spot that a mandatory field is empty, that a check has been skipped or that a result has not yet been reviewed.

What these four share is their form: at a fixed point in the work process, the system compares this patient's data with a set of rules and shows the outcome to the person acting at that moment. What they do not share is the risk involved, and it is precisely that difference that later determines both the design and the regulatory framework.

It also works slightly differently everywhere. In a hospital, most of the gain lies in the moment of prescribing and in following up results. In a GP practice, it is more often about repeat prescriptions, review appointments that lapse and spotting patients who drop out of view. In long-term care, it is about agreements that must hold across many staff and many shifts, where the person who sees the alert is the care worker rather than the doctor. In mental health care, the emphasis is more often on monitoring agreed steps in a treatment pathway than on lab values. These differences determine who the user is, and therefore what an alert should look like and how much can be asked of someone at the moment it appears.

It is just as important to be clear about what it is not. Decision support does not make a diagnosis, does not take a treatment decision and does not replace professional judgement. Nor do we claim any effect on care outcomes on this page: what an alert achieves depends on the content of the rule, on the moment it appears and on the clinician who works with it. A supplier that does promise you such an effect is promising something that cannot be built into software.

Are you in the right place? This page is about the alerting layer itself. If you are looking for a broader care system, with record keeping, scheduling, billing or a client portal, start with custom healthcare software development. If your question is mainly about demonstrably meeting requirements, with logging, authorisations and audit trails at its core, then healthcare compliance software is the better starting point.

The clinician decides, not the software

The software does not make a diagnosis or take a treatment decision. It shows what is known and why it stands out right now. The judgement stays with the professional who signs off.

Timing determines the value

The same alert is useful while prescribing and useless in an overview for next week. Decision support is therefore above all a question of timing.

No source, no signal

What the system cannot read from the record in real time, it cannot assess either. Here the integration is not a precondition but half of the product.

Why most alerts get clicked away

Almost every decision support project gets stuck at the same point sooner or later: alerts appear, and those alerts get clicked away. Not out of unwillingness and not out of incompetence, but because the behaviour makes sense given what the system shows. A rule that warns about every theoretically possible interaction mostly warns in situations the prescriber has known about for ages, knowingly accepts and is already monitoring. Anyone who receives an alert a hundred times that does not apply learns to dismiss it before reading it. From that moment on, the hundred-and-first, the one that does matter, gets exactly as much attention as all the rest.

The cause rarely lies in the knowledge source and almost always in the lack of context. A generic rule knows that drug A and drug B call for attention when taken together. What the rule does not know is that this patient has been taking that combination for months, that the indication explains the combination, that the dosage has been adjusted, that the check the alert asks for was done last week, or that a colleague already assessed the same alert an hour earlier. All of that information is in the record. It is just not in the rule.

On top of that, dismissing an alert usually goes nowhere. In many systems an ignored alert disappears without a trace, at most with a counter behind it. Yet that is the most useful information in the entire system: every override is a clinician telling you that the rule is wrong here. If that reason is recorded, categorised and read periodically, you have a plan for refining the rules. If it is not, every discussion about alerts comes down to a matter of opinion.

The content of the alert itself matters too, and it is often overlooked. A useful alert tells you at a glance what is going on, why that conclusion was drawn and which data from this record it is based on, with the source of the rule included for anyone who wants to look it up. It also offers the actions that logically follow, so the clinician does not first have to go three screens further to do something with the information. And when an alert is rejected, it asks for a reason from a short, meaningful list, not an empty text field that everyone skips. An alert that merely says something is wrong shifts the work to the user; an alert that shows what it is based on gives that user something to judge by. Here too the division of roles remains unchanged: the software provides the evidence, the clinician draws the conclusion.

From more alerts to fewer, weightier ones

The way out is not smarter warnings at the same volume, but fewer warnings. That requires distinguishing between levels of severity. A small share of the rules justify an interruption that the user cannot get past without giving a reason. The vast majority should be passive: a marker on the screen, a line in an overview, a colour next to a value, something that stands out without halting the work. Anything that fits neither category does not belong in the workflow but in a periodic overview for whoever reviews the record, such as the pharmacist or the quality officer.

Suppression based on the record also helps: an alert that has already been assessed for this patient and this prescription does not return as long as nothing changes, and an alert that is routinely dismissed at ward level goes onto a review list instead of onto the screen. That is not a technical trick but a clinical choice, and therefore one made by clinical expertise, not by the developer. Measurement is part of it: for each rule, how often it fires, for how many different patients, how often it is acted on, and for what reason it is dismissed. Without those figures from your own system, any statement about its effect is an assumption.

What a decision support system needs to be able to do

These components are interdependent. Monitoring without context produces noise, rules without management become outdated, and logging without feedback turns a log file into an archive nobody reads. In all six cases the same applies: the output is information for the healthcare professional, who decides for themselves what happens.

Medication monitoring with context

Checks for interactions, duplicate medication, dosage and contraindications, where the known data for this patient suppresses the alert or makes it more severe.

Protocol at the moment of action

The step from the guideline appears on the screen where the work is done, with room to deviate with a justification and with the source alongside.

Flagging abnormal values

Ranges per patient group, the trend alongside the individual measurement, and a route to the person who can act on it at that moment.

Record checks and completeness

Missing fields, unreviewed results and skipped checks, presented as a worklist rather than as an interruption.

Rule management with versioning

Every rule has an owner, a traceable source, an effective date and a history, so you can see afterwards what applied at any given moment.

Logging what was shown

Which alert, to whom, at what moment, what was chosen and for what reason. This is the basis for accountability and for refining rules.

The integration with the source system determines whether it works

Decision support is not the hardest part of such a project in terms of content. The rules are usually described somewhere, in a guideline, a formulary or a local working agreement. Where it falls apart is whether the system has the right data at the right moment, and whether the answer ends up on the screen where someone is looking at that moment. That is an integration question, and in practice it determines whether you are building a useful tool or a second screen nobody opens.

There are roughly three levels. The first is a periodic export from the source system: easy to implement, but by definition after the fact, so only useful for overviews and retrospective checks. The second is event-driven: the source system sends a message when there is a new prescription, a new result or an admission, and the decision layer responds to it. That works for flagging, but the healthcare professional has to see the answer somewhere else. The third is the only level at which genuine decision support emerges: at the moment of action, the source system calls a service and displays the answer in its own screen. Then the timing is determined not by your application, but by the work process.

Technically, this often runs over HL7 version 2 messaging, which is still widely used in Dutch healthcare institutions, or over FHIR when the source system offers it. For calling the logic from within the workflow, CDS Hooks is an open pattern designed for exactly that purpose. More important than the choice itself is whether your supplier permits this type of integration, which data is available and whether you are allowed to write anything back. A read-only integration without write-back means the assessment of an alert ends up outside the record, and therefore cannot be found by the next clinician.

Two things are consistently underestimated. The first is coding: as long as allergies, indications or laboratory results arrive as free text, no rule can rely on them. Coded data, using a shared terminology such as SNOMED CT for clinical concepts and LOINC for laboratory tests, is a prerequisite, not a bonus. The second is timing: a call made while prescribing has a hard limit within which the response must arrive, otherwise the user waits or the alert disappears unnoticed. A missing response should therefore itself be visible, because a silent failure in a safety function is more dangerous than a failure that everyone can see.

Two further things belong in the integration design rather than at delivery. The first is authorisation: an alert contains patient data, so who gets to see it follows from the treatment relationship and the role, not from the fact that someone happens to have the screen open. The second is minimisation: retrieve what the rule needs, not the entire record, because every extra field you copy becomes a field you have to secure, retain and clean up. For logging, the opposite applies: for the alert itself you should record plenty, namely what was shown, to whom, on what basis and what was chosen. That is precisely the information you need when reviewing an incident, and also the information that makes it possible to refine rules.

  • Current data, not last night's export
  • A call at the moment of action, not afterwards
  • Coded data instead of free text
  • Writing back what was shown and chosen
  • A hard time limit for the response
  • Failing visibly: no response is also an alert
  • Agreements with the supplier of the source system
Not sure about a large project yet?

Test your idea first: a working prototype in 1 day

With OneDayBuild, we make your idea tangible in a single day for €1,150, so you know whether further development is worth the investment. Decide to go ahead with the full build? Then we deduct the cost in full.

View OneDayBuild →

Who maintains the rules when the guideline changes

A set of rules is not a deliverable but a living file. Guidelines are revised, formularies are adjusted, a drug disappears from the market, a local working agreement becomes national or the other way round. A system that is not set up for this ages without anyone noticing, and that is the most troublesome form of ageing: the alerts keep coming, they are just no longer correct. The question of who maintains the rules therefore belongs in the design, not in a maintenance contract afterwards.

The first design decision is to keep the rules out of the code. If the content of a rule sits in the software, every change to a guideline requires a software release, and a developer becomes the gatekeeper of clinical content. That is an uncomfortable role for both parties. A workable approach is to separate the engine, which retrieves data, evaluates conditions and presents the result, from the content, which is held in a maintainable form that the clinical owner can access. What exactly that form looks like depends on who does the maintenance; a pharmacist with the time and the aptitude can handle more than a committee that meets twice a year.

The second decision is ownership per rule. Not per system, per rule: a name, a discipline, a source the rule comes from with the version included, an effective date and a moment when it is reviewed again. If the source changes, every rule that refers to it appears on a list. Without that reference, every review is manual work in which someone has to remember from memory which rules are affected.

The third decision concerns introducing rules. A new or changed rule should first run without being visible, against real data, so that it is clear in advance how often it would fire and for which type of patient. A rule that fires dozens of times a day in that shadow run is not ready. In addition, for every change it should be recorded who proposed it, who approved it and from when it applied. That last point is not bureaucracy: when reviewing an incident, the only sensible question is which rule version was active at that moment, and therefore what information the healthcare professional did or did not have in front of them. Here too the division of roles does not change: the system records what it showed, and the healthcare professional remains the one who made the decision.

Finally, a practical warning for the tender or the project budget. Self-built decision support rarely goes wrong during the build. It goes wrong the year after, when the person who knew the rules has moved on to something else and nobody has been formally designated to maintain them. Settle that before you start. If the protocol spans multiple steps and practitioners, this directly affects how you set up care pathway software; if the emphasis is on medication, it affects pharmacy software.

HL7 version 2 FHIR CDS Hooks SNOMED CT LOINC Rule version control Shadow running Override management Audit trail

Medical device or not, and what that means

This is the part where you need a specialist, and where we therefore only describe what is stated in public government sources. Software with a medical purpose can fall under the European rules for medical devices. The Dutch central government summarises the condition as follows: software is a medical device if the manufacturer states that the software directly controls a medical device, provides information about patient care that leads to better decisions, or supports the treatment of the patient. By definition, decision support is close to the last two descriptions. That is not a side detail of your project, it is a design starting point.

The framework is the European Medical Device Regulation, Regulation (EU) 2017/745, known as the MDR, which according to the Dutch Health and Youth Care Inspectorate (IGJ) became applicable on 26 May 2021. The law has four risk classes: class I as the lowest, then IIa, IIb and III as the highest. The manufacturer is itself responsible for correctly determining that class. For the content of this page, the following matters most, again in the wording of the Dutch government: software that provides information for decisions with diagnostic or therapeutic purposes falls into class IIa, for example, and if decisions based on that information can lead to a deterioration in someone's state of health, the software falls into the higher risk class IIb. Stricter requirements apply to higher classes in the CE marking procedure, and an inspection body, the notified body, assesses the product and either does or does not issue an MDR certificate for it. See the Dutch government's explanation and the IGJ's supervision.

For the build, that means three things. There are requirements for the product itself: Annex I of the MDR requires, among other things, that software is developed in line with the state of the art, taking risks and information security into account. There are requirements for the documentation around it, including risk analysis and clinical evaluation; the IGJ (the Dutch Health and Youth Care Inspectorate) names exactly those two as recurring weak points among manufacturers. And there is supervision: the IGJ monitors compliance and can enforce. The practical consequence for your planning: build that documentation alongside the software from the start, because reconstructing afterwards what was weighed up during development is considerably more expensive than recording it at the time.

A separate route exists for healthcare providers that make and use devices in-house. In its intervention policy, the IGJ states that the requirements of the MDR and the IVDR do not apply when in-house manufactured medical devices meet the requirements of Article 5(5) of those regulations. This is explicitly not a back door: the article sets conditions, and the route covers manufacture and use within your own institution. If you want to rely on it, have it checked beforehand that you really fit within it.

If your decision support works with a model that learns from data rather than with fixed rules, the European framework for artificial intelligence applies as well. That is Regulation (EU) 2024/1689, which according to the European Commission entered into force on 1 August 2024 and became applicable on 2 August 2026, with an extended transition period until 2 August 2028 for high-risk AI built into products that already fall under European product legislation. You can read more at AI Act compliance software and on the European Commission's page.

This is not legal advice. Whether your software qualifies as a medical device, and in which class, depends on the purpose you define as manufacturer and on how the software is used. Settle that with a notified body or a specialist in medical device regulation before you build, not afterwards. What we add at the table is the translation into design: which part of the desired functionality drives up the risk, and whether that part really needs to be included to solve the problem.

Building alongside the care system, or inside it

The first question is not whether you build, but what your current supplier already offers. Medication monitoring in a prescribing system is a mature product with a maintained knowledge base behind it. Rebuilding that is rarely wise, if only because you would then also be maintaining the knowledge base. The same applies to standard checks that every institution needs.

Building alongside the existing system becomes interesting in four situations. When the rule is specific to your organisation or population and no supplier is going to build it. When you need to combine data from several systems and the source system does not know the other source. When you want to adjust faster than your supplier's release calendar allows. And when the alerting spans organisational boundaries, so that none of the systems involved sees the whole picture.

Building inside, that is, configuring within the existing system, is better when that system has a usable rules engine, and certainly when the alert has to appear on a screen you cannot reach from outside. That last point is more often decisive than the functional comparison. Either way, keep the scope small: the narrower the medical purpose you define, the lighter the process that follows from it. A worklist for the pharmacist has different consequences from an interruption in the prescribing screen, even if the underlying rule is identical. And in both cases the division of roles stays the same: the software alerts, the healthcare professional makes the diagnosis and takes the treatment decision.

  • Define the intended purpose before you choose functionality
  • The manufacturer determines the risk class and is responsible for it
  • A higher class means a notified body is involved in the process
  • Build in risk analysis and clinical evaluation from the start
  • In-house manufacture follows its own route with its own conditions
  • If the model learns from data, the AI framework applies as well
  • Have a specialist review the qualification and class

Frequently asked questions about clinical decision support

No. The software does not make a diagnosis and does not take treatment decisions. It brings information to attention at the right moment: a possible interaction, a step from a protocol, a value outside the agreed range or a record that is incomplete. The healthcare professional assesses that information, weighs it against everything else they know about this patient, and decides. The system records what it showed and what was chosen, so it can be traced afterwards what information was available. Nor do we claim any effect on care outcomes: what an alert achieves depends on the content of the rule, on the moment it appears and on the professional who works with it.

It can be, and with decision support that question is rarely theoretical. The Dutch government describes software as a medical device when the manufacturer states that the software directly controls a medical device, provides information about patient care that leads to better decisions, or supports the treatment of the patient. The framework is the European Medical Device Regulation, Regulation (EU) 2017/745, which has applied since 26 May 2021 according to the IGJ (the Dutch Health and Youth Care Inspectorate). There are four risk classes and the manufacturer decides which one applies. Software that provides information for decisions with diagnostic or therapeutic purposes, for example, falls into class IIa, and into the higher class IIb where those decisions could lead to a deterioration in someone's state of health. This is not legal advice: have the qualification and class reviewed by a notified body or a specialist.

By creating fewer of them and supporting them better. An alert should take into account the data from the record that the healthcare professional already knows: has this patient been taking the combination for some time, does the indication explain it, was the requested check done last week, has a colleague already reviewed the same alert. Distinguishing by severity also helps: a small share of the rules justify an interruption, the vast majority should be passive and the rest belong in a worklist for whoever reviews the record. Then measure for each rule how often it fires, how often it is acted on and the reason given when it is dismissed. Every dismissal is a healthcare professional telling you the rule does not fit here.

That depends mainly on what your supplier allows. There are three levels. A periodic export is simple but always after the fact, so it is only useful for overviews. Event-driven messaging, often using HL7 version 2 or via FHIR, is near real time but shows the answer outside the screen where the work is being done. Only at the third level does genuine decision support arise: the source system calls a service at the moment of action and displays the answer itself, a pattern for which CDS Hooks provides an open standard. The prerequisites are coded data instead of free text, a hard time limit for the response and the ability to write back what was shown and chosen.

Someone with a name and a title, appointed before the system goes live. In practice this means: the content of the rules does not sit inside the software but in a manageable form, every rule has an owner and a traceable source with a version, and if that source changes, every rule that refers to it appears on a review list. A changed rule first runs silently, without being visible, against real data, so you know in advance how often it would fire. For every change, a record is kept of who proposed it, who approved it and from when it applied, because when reviewing an incident the question is which version of the rule was active at the time. Self-built decision support rarely goes wrong during the build, but in the year after.

First look at what your current supplier already offers. Medication monitoring in a prescribing system is a mature product with a maintained knowledge base behind it, and rebuilding that means taking on that knowledge base as well. Building your own alongside the existing system becomes interesting if the rule is specific to your organisation or population, if you need to combine data from multiple systems, if you want to adjust faster than the release calendar allows, or if the alerting crosses organisational boundaries. Configuring within the existing system is better when that system has a usable rule engine or when the alert has to appear in a screen you cannot reach from outside. In both cases, keep the scope small.

Related services

Building healthcare software

If your question goes beyond the alerting layer, with record keeping, planning, billing or a client portal, start with building healthcare software. That covers the broader picture; this page is only about the decision support itself.

Healthcare compliance software

If the question is about demonstrably meeting requirements, with logging, authorisations, retention periods and audit trails at its core, healthcare compliance software is a better fit than a decision layer.

Patient intake software

Many checks start with what was or was not asked during intake. If data is missing there, no rule can build on it. See patient intake software.

Start with the alerts you are missing today

Tell us which three signals you would want today, which system the data for them comes from, and who will manage the rules afterwards. Those three answers usually show whether this is configuration in your current system, a separate layer alongside it, or first a conversation about qualification as a medical device. The same starting point applies in all three cases: the software alerts, the healthcare professional makes the diagnosis and takes the treatment decision.

Edit content