Categories
News

Artificial intelligence methods applied to longitudinal data from electronic health records for prediction of most cancers: a scoping review | BMC Medical Research Methodology


Retrieved research

Looking the 6 databases returned 1214 research, of which 414 have been duplicates. An extra 61 research have been retrieved from reference and quotation searches. Following screening and eligibility evaluation, 35 research have been included within the remaining review. A flowchart displaying the choice course of is offered in Fig. 2. The quantity of research revealed by yr is proven in Fig. 3.

Fig. 2
figure 2

Circulate diagram for examine identification and choice. Developed utilizing the PRISMA template offered in [21]

Fig. 3
figure 3

Quantity of retrieved research by yr of publication. *12 months to date 09/08/2024

Research traits

Research setting and inhabitants

Research used populations from the USA (19, 54%), the Netherlands (4, 11%), Taiwan (5, 14%), Denmark (2, 6%), Sweden (1, 3%), South Korea (1, 3%), Israel (1, 3%), and Singapore (1, 3%). One examine didn’t report the place the inhabitants originated from, and one examine used an extra dataset from the UK as a validation set. 5 research (14%) used single-centre data for mannequin growth, 20 (57%) used data from a number of centres linked by location or healthcare supplier, 9 research (25%) used nationwide datasets, and one examine (3%) didn’t report the examine setting. The nationwide research originated from Sweden, Taiwan, South Korea, and Denmark, whereas the multi-centre research utilizing data from affiliated practices originated within the USA (n = 15), the Netherlands (n = 4), and Israel (n = 1). The research used case–management (9, 26%), nested case–management (6, 17%), or cohort (20, 57%) examine design. Of the settings for the datasets used, 4 research used major care (11%), seven used secondary care (20%), 23 used each major and secondary care data (66%), and one didn’t report.

Outcomes/Prediction Process

Ten research predicted the danger of most cancers inside a particular time frame. Twenty research centered on both detection or early detection of most cancers. One examine predicted metastasis and 4 research predicted recurrence.

The commonest cancers included within the research have been pancreatic and colorectal most cancers (each 9 research, 26%). There have been 6 research predicting lung most cancers (17%), 3 research (9%) every contemplating liver and gastric most cancers and a pair of (6%) contemplating breast, pores and skin, leukaemia, and oesophageal most cancers respectively. Mind metastasis, most cancers of the small gut, anal most cancers, cervical most cancers, and prostate most cancers have been every predicted in a single examine respectively. Moreover, one examine predicted most cancers as a generic consequence, with no website specified. Word that some research developed fashions for a number of websites.

Scientific options

Probably the most generally included options have been demographics (25, 71%), diagnoses (22, 63%), laboratory checks (22, 63%), and prescriptions (18, 51%). Different options included signs, referrals, procedures, free textual content notes, life-style components, photographs, tumour staging, and histological options. Frequency of the options is proven in Fig. 4.

Fig. 4
figure 4

Frequency of research utilizing every kind of scientific function, the place the whole quantity of research is 35. The class ‘Different’ contains research utilizing photographs, histology options, and tumour staging

The scientific variables chosen various between the approaches taken. All of the research utilizing function engineering used laboratory checks of their fashions, whereas solely a third of the fashions utilizing sequential inputs did the identical. As well as, all however one of the function engineering fashions used demographic data, in distinction to round two-thirds (n=12) ofthe sequential enter fashions.

Mannequin traits

Methods for representing temporal info inside predictive fashions will be divided into two major approaches. The primary method utilises function engineering, a course of the place data are extracted and manipulated to kind informative variables, to discover significant representations of temporal data or seize key temporal traits. These options, generated for every affected person, can be utilized as inputs in downstream AI fashions to generate a person classification or threat rating. For this method, we analyse the methods for representing the longitudinal info, somewhat than the following AI mannequin as these are usually circuitously tailor-made to deal with longitudinal data. The second method makes use of temporal sequences as direct inputs. This method usually contains a mannequin that has particular mechanisms to mannequin the sequential nature, particularly the dependence between time steps.

Function engineering for illustration of sequential data

Sixteen fashions used function engineering for the illustration of temporal data. Numerous approaches have been taken; these are summarised in Desk 1.

Desk 1 Function engineering methods utilized by research within the review

Three research used ‘pattern’ or ‘slope’ options [26,27,28]. 5 used absolutely the change between outlined time factors, for instance, the 12 month change earlier than prediction date [29,30,31,32,33]. For 2 of these research [29, 30], the values used to calculate absolutely the change have been inferred from fashions educated on every sufferers particular person trajectory for that variable. Kinar et al. [29] match linear regression fashions which have been used to predict values at 18 and 36 months earlier than to the index, and the change between these time factors was used because the pattern function. The same method was utilized by Rodriguez et al. [30], becoming a linear combined results mannequin to log-transformed laboratory measurements to present a ‘smoothed’ trajectory (i.e., with decreased measurement error). Predictions have been then used to calculate the 6-month change in log-transformed measurement.

Rubenstein et al. used the biggest enhance and whole variation of a measurement as options [34]. Labaratory dynamics have been utilized by Beinecke et. al. [35] One examine acknowledged ‘trending’ options have been used, however didn’t describe how these have been calculated [36]. In three research, sample mining was used to discover predictive temporal patterns [37,38,39]. One examine applied Wavelet and Fourier transforms to longitudinal data and used the coefficients as options to a mannequin [40].

Unsupervised approaches have been additionally used to extract options from time-series. Lasko et al. and Ho et al. each use autoencoders to study normal representations of a affected person’s trajectory [40, 41].

Fashions taking sequential data as direct enter

Twenty research used deep studying fashions with a sequential enter, both the uncooked sequence or ‘binned’ into discrete time intervals. The methods used are summarised in Desk 2.

Desk 2 Abstract of methods utilized by research within the review which take a sequential enter

Ten research used fashions based mostly on recurrent neural networks (RNNs): seven used lengthy Quick-Time period reminiscence fashions (LSTMs) [40, 46,47,48,49,50,51] and 7 used gated recurrent items (GRUs) [40, 46, 49, 50, 52,53,54,55]. Two research [53, 54] used the reverse time consideration mannequin (RETAIN) proposed by Choi et al. which introduces an consideration mechanism to a GRU to prioritise essentially the most significant visits in a affected person’s enter sequence [56].

5 research used convolutional neural networks (CNNs) [40, 57,58,59,60], whereas one examine [61] used a CNN-LSTM, representing diagnoses and medicines as a 2D matrix and performing 2D convolutions over the enter.

Three research [62,63,64] used a commonplace feed-forward neural community, the place every time-step was represented by a node within the structure. In two of these, Park et al [63, 64] educated a separate neural community for every variable as an ‘embedding community’ to scale back the dimensionality of the enter, and these decreased options have been concatenated to kind an enter to a remaining classification community.

Six research utilised transformer architectures [40, 49, 54, 55, 61, 65]. Positional encodings have been derived in a single examine utilizing the widespread method of evaluating sinusoidal capabilities of various frequencies on the level the token seems within the enter sequence [40]. Two research adjusted this method in order that the sinusoidal capabilities have been evaluated at a affected person’s age, somewhat than the place inside the sequence [49, 55]. Rasmy et al. introduce multi-layered embeddings for place, denoting not solely the order of visits, but additionally the order of codes inside the visits [54]. Two research didn’t report the tactic of place embedding [61, 65].

The deep studying methods used require inputs of uniform size. There have been a quantity of approaches to addressing lacking data alongside the temporal axis. 5 research had categorical options representing the presence of an occasion inside a particular window, therefore the size of inputs didn’t want particular consideration [51, 58,59,60,61]. The place occasions are represented as embeddings or numerical values are used, any sequence that’s shorter than the utmost sequence size have to be coerced not directly. 5 research [40, 47, 48, 50, 65] ‘padded’ the enter by including zero vectors to the sequence. Three research [46, 52, 57] used forward-filling, the place lacking data alongside the temporal axis is stuffed through the use of the latest current measurement. Six research didn’t report the tactic of addressing enter sequence size [54, 55, 62,63,64].

Prediction home windows

The prediction home windows for every mannequin, as outlined in 2.4, are proven in Tables 3 and 4. Desk 3 reveals the time home windows utilized in every of the danger prediction fashions. For the commentary window, all however one of the danger prediction research used the total obtainable data inside the examine interval and didn’t impose any restrict on data earlier than the index date. The prediction home windows various between 3 and 60 months, with 36 months being the most typical. Just one of the danger prediction research investigated a number of prediction home windows [55]. This examine introduced the bulk of their outcomes with respect to the 36-month prediction window, stating that it’s a cheap window for screening.

Desk 3 Time home windows utilized in threat prediction fashions
Desk 4 Time home windows utilized in every of the most cancers detection fashions

Desk 4 reveals the time home windows for the 20 most cancers detection fashions. 5 of these used the total historical past of the affected person because the commentary window. The commentary home windows of these research that restricted the window ranged between 6 and 60 months. Two research didn’t report their commentary window. One examine didn’t outline their commentary window by time interval, however somewhat by quantity of measurements. 9 of the research haven’t any lead time, as an alternative detecting most cancers utilizing all data that was obtainable earlier than most cancers analysis. Eleven research investigated early detection of most cancers utilizing a lead time, these various between 3 and 36 months. 5 research investigated quite a few lead occasions. Observe-up time was poorly reported in most research, with 14 research not offering info on this window. 4 research didn’t follow-up controls for analysis, whereas two research gave a follow-up time of 36 months.

The home windows for metastasis and recurrence prediction fashions are proven in Desk 5. One examine included comply with up of controls [53]. Two research outlined the beginning of the commentary window as a particular scientific occasion relating to the first most cancers [35, 40].

Desk 5 Time home windows utilized in metastasis or recurrence prediction fashions

Comparability to cross-sectional fashions

Of the research included on this review, seven in contrast longitudinal methods to cross-sectional approaches, utilizing data from solely a single timepoint [26, 37,38,39, 52]. Ioannou et al. reported enchancment in discrimination and calibration over cross-sectional outcomes when utilizing the sequential enter mannequin, however no important enchancment utilizing engineered options [52]. Kop et al. present in one examine that engineered temporal options improved predictions [37], however this end result was not repeated in later work [38]. Hoogendoorn et al. didn’t report an enchancment in predictive efficiency, however did observe that efficiency was extra secure throughout totally different data sorts in sensitivity evaluation [39]. Learn et al. famous a pattern in direction of enchancment, however the outcomes weren’t definitive [26]. Within the multimodal examine by Li et al., longitudinal was discovered to enhance predictions in modalities integrating each picture data and scientific data, however not in all modalities [65].

Explainability

Twenty-two research thought-about both explainability of predictions or mannequin reasoning. 13 of these (10 function engineering fashions, 3 sequential enter fashions) introduced mannequin stage interpretability reminiscent of function importances to display what info is utilized by the mannequin to make predictions. 9 research (1 function engineering mannequin, 8 sequential enter fashions) had prediction stage explanations, the place the components contributing to a person prediction are calculated. The methods used for particular person prediction explanations have been native interpretable model-agnostic explanations (LIME) [47], attention-based interpretation [49, 53, 54], built-in gradients [55, 61], and Shapley additive explanations (SHAP) [34, 36, 53, 64].

Reproducibility of analysis

13 research (36%) had code obtainable to use on-line. Two research used data that was tailored to a widespread data mannequin: Kim et al. used the Observational Medical Outcomes Partnership Frequent Data Mannequin (OMOP-CDM) [47], and Jia et al. [28] used data adhering to the TriNetX commonplace data mannequin [66]. One examine [47] used data which is freely obtainable on-line. Fifteen research used datasets which will be requested or bought: the Veterans’ Affairs Company Data Warehouse [31, 32, 34, 52, 55], the Kaiser Permanente Southern California databank [30,31,32,33, 36], the Julius Basic Practitioner Community [38, 39, 46], Cerner Health Info [53, 54], HCUP State Inpatient Databases (SID) [50], IQVIA datasets [51], and TriNetX [28]. Six research used data that’s obtainable to researchers inside the nation of origin solely [27, 55, 58, 59, 61, 62].

High quality Evaluation

An summary of the area judgements for the PROBAST evaluation are proven in Fig. 5 and particular person judgements are offered in Extra File 2. The general threat of bias was excessive for 90% of the research within the review, low for 7.5%, and unclear for 2.5%.

Fig. 5
figure 5

A abstract of threat of bias judgements assessed utilizing the PROBAST framework. Word that some research could have a number of threat of bias assessments the place exterior validation was carried out or the examine included a couple of predictive mannequin



Source link

Leave a Reply

Your email address will not be published. Required fields are marked *