SET Domain

The revised FIQ (the FIQR) has 21 individual questions (Table 1). All questions are based on an 11-point numeric rating scale of 0 to 10, with 10 being 'worst'. As in the FIQ, all questions are framed in the context of the past 7 days. Following the convention used in the FIQ, the FIQR is divided into three linked sets of domains: (a) 'function' (contains 9 questions versus 11 in the FIQ), (b) 'overall impact' (contains 2 questions, as in the FIQ) but the questions now relate to the overall impact of FM on functioning and the overall impact symptom severity, and (c) 'symptoms' (contains 10 questions versus 7 in the FIQ); one original FIQ symptom was dropped: 'When you worked, how much did pain or other symptoms of your fibromyalgia interfere with your ability to do your work, including housework?' The symptom domain contains four new questions relating to memory, tenderness, balance, and environmental sensitivity (to loud noises, bright lights, odors, and cold temperatures). The 'time' dimension is the same as the FIQ; that is, all questions relate to the impact of FM over the course of the past 7 days. The scoring of the FIQR is much simpler than the FIQ: namely, the summed score for function (range 0 to 90) is divided by 3, the summed score for overall impact (range 0 to 20) is not changed, and the summed score for symptoms (range 0 to 100) is divided by 2. The total FIQR is the sum of the three modified domain scores. The weighting of these three domains is different from the FIQ in that 30% of the total score is ascribed to 'function' as opposed to 10% in the FIQ, 50% is ascribed to 'symptoms' as opposed to 70% in the FIQ, and 'overall impact' remains the same as the FIQ at 20%. The total maximal score of the FIQR remains the same as the FIQ, namely 100.

The PROM/PREM implemented in this project were those proposed in the PCB set: questionnaires at two moments during pregnancy (T1: first trimester, T2: early third trimester) and three postpartum (T3: maternity week, T4: 6 weeks postpartum, T5: 6 months postpartum). The PCB set was developed internationally and subsequently translated to the Dutch setting, both phases involving all stakeholders, including care professionals and patients [18 (link), 29 (link)]. An overview of the PCB set’s patient-reported domains and timeline for completion is provided in Additional file 1: Fig. S1. The set’s PROM/PREM were implemented for two purposes. First, individual-level PROM/PREM were implemented in clinic: reviewing N = 1 scores with patients during a regular care contact after completing a questionnaire. The timeline of collection, workflow, and follow-up services (including scoring and alert values) were organized as described in the national pilot project [30 (link)]. Second, the same PROM/PREM outcomes would be used at group-level in network-broad QI sessions. Despite the complexity of combining these purposes, findings in our pre-implementation research amongst care professionals, patients and other stakeholders in perinatal care suggested both goals could also reinforce each other [8 (link)]. Direct usability in clinical practice could, for instance, motivate care professionals and patients to comply, thereby generating data for group-level use (and vice-versa). Likewise, other previous findings from our pre-implementation analysis and feasibility pilot [8 (link), 28 (link)], were used to design the initial implementation strategy. Important elements for individual-level use included visual alerts to support care professionals in interpreting the answers and offering patients a choice whether their care professional had insight in their individual PREM answers. During the action research project, this initial implementation strategy (Fig. 1) was continuously refined guided by action research principles in iterative cycles of planning and executing implementation activities, data generation, and reflection on these data to refine subsequent activities. These cycles were conducted jointly by researchers and care professionals. The researchers developed the baseline strategy for project organization and education (e.g. identified possible IT-systems, developed an e-learning and kick-off meeting), provided materials and support for its execution (e.g. patient information folder, for working protocol for care professionals), and facilitated data generation for its refinement (e.g. organized focus groups, sent out the survey). The project teams designed and coordinated local implementation (e.g. adapt instruction material to local workflow, chose the IT system that best fitted local needs and resources) and participated in data generation and reflections (e.g. survey results were discussed in project team meetings, participation in focus groups). Three OCN started implementation sequentially to be able to learn from previous experiences, exchanged via the researchers and directly between care professionals from different OCN. After the one-year implementation period, project teams reported their experiences to their OCN and advised future steps in an end-evaluation.Fig. 1

Timeline of implementation and data generation activities. PROM, patient-reported outcome measure. PREM, patient-reported experience measure. QI, quality improvement. OCN, obstetric care network. CP, care professional. VHBC, value-based healthcare

Content analysis was performed to assess the impact of the H&P 360 template. For purposes of qualitative comparison of note content, research coordinators collected and deidentified all of the H&P 360 notes and a sample of the traditional notes. The sample of traditional notes was drawn by attempting relatively balanced representation across students. Specifically, each student could contribute no more than 5 traditional notes to the total sample; for those with more than 5 traditional notes, a random subsample was selected for inclusion.
The content analysis team was composed of three internists involved in medical education (JWT, VGP, IJA) and one medical student (EYR). Throughout the process of analysis, team members discussed their preconceived notions and biases from their roles in education and patient care. The team began with a set of a priori content domains based on the H&P 360 template (eg, mental health, behavioral health, social support). The team members independently reviewed a set of notes—four from the H&P 360 group and four from the traditional group—to clarify the definition of the content domains, add additional de novo content domains as needed, and improve consistency between coders. Subsequently, for each of the notes, two team members extracted relevant text and entered it into a Research Electronic Data Capture (REDCap) template under the appropriate content domain. Discrepancies in coding were reviewed and resolved through email correspondence. The text from each content domain was then aggregated into a document and reviewed by two members of the team to identify themes within each content domain and to assess whether there were qualitative differences in the content between the H&P 360 and traditional templated notes. Each content domain was discussed at the weekly group video meeting. The number of notes categorized under each content domain was counted for the H&P 360 and traditional templated groups.

We define transfer learning as in the following clear and simple statements:
"Transfer learning and domain adaptation refer to the situation where what has been learned in one setting (e.g., distributionP₁) is exploited to improve generalization in another setting (say, distributionP₂)”⁴⁹ ;
"Given a source domainD_Sand learning taskT_S, a target domainD_Tand learning taskT_T, transfer learning aims to help improve the learning of the target predictive functionf_T( ⋅ ) in D_Tusing the knowledge inD_SandT_S, whereD_S ≠ D_TandT_S ≠ T_T”^{50 (link)}.
In our study the source and the target tasks are the same, i.e. T_S ≡ T_T. The task is always to perform sleep staging with the same set of sleep classes/stages. We want to transfer the knowledge about the previously learned sleep recordings (e.g., different hardware, different subject distributions with different sleep disorders) and the knowledge about the sleep scoring-rules (i.e., inter-scorer variability in the different data centers). The process generally involves overwriting a knowledge from a small-sized database to a previous big-sized knowledge (result of a long training process). One big concern is to avoid ending up in what the data scientists call catastrophic forgetting: “Also known as catastrophic interference, it is the tendency of an artificial neural network to completely and abruptly forget previously learned information upon learning new information” as defined in⁵¹ . Even if it is conceptually easy to understand, avoiding its occurrence is not trivial. To partially bypass this phenomena we fine-tune the architecture on the target domain using a smaller learning rate.
In our experiments we first pre-train the architecture on the data-source domain S (e.g., a set of different domains/databases $\left\{S_{{\mathrm{DB}}_1},S_{{\mathrm{DB}}_2},...,S_{{\mathrm{DB}}_{\mathrm{n}}}\right\}$ ), then we fine-tune the model on the data-target domain T. Formally, we first minimize the loss function L_S, resulting in the learned parameters θ: $argmi{n}_{\theta }=\sum _{\left(\mathbf{x},\mathbf{y}\right)\in D_{\mathrm{S}}}^{}{L}_{\mathrm{S}}\left(\mathbf{x},P\left(\mathbf{y}\parallel \mathbf{x}\right),P_{\theta }\left(\mathbf{y},\mathbf{x}\right)\right)$
The parameters θ of the pre-trained model are used as the starting point on the data-target domain T. To transfer the learning on the new domain T, we fine-tune all the pre-trained parameters ${\theta }^{\prime }=\theta$ (i.e., the entire network is further trained on the new data domain T): $argmi{n}_{{\theta }^{\prime }=\theta }=\sum _{\left(\mathbf{x},\mathbf{y}\right)\in D_{\mathrm{T}}}^{}{L}_{\mathrm{T}}\left(\mathbf{x},P\left(\mathbf{y}\parallel \mathbf{x}\right),P_{\theta }\left(\mathbf{y},\mathbf{x}\right)\right)$