Schlüsselwörter Kolposkopie - maschinelles Lernen - rekursive Partitionierung - zervikale Dysplasie
Keywords colposcopy - machine learning - recursive partitioning - cervical dysplasia
Abbreviations
AGC:
atypical glandular cells
AIC:
Akaike information criterion
ASC-H:
atypical squamous cells, cannot exclude HSIL
AUC:
area under the curve
LSIL:
low-grade squamous intraepithelial lesion
HPV:
human papillomavirus
HSIL:
high-grade squamous intraepithelial lesion
LEEP:
Loop Electrical Excision Procedure
PCR:
polymerase chain reaction
ROC:
receiver operating characteristics
SD:
standard deviation
Introduction
Infection with human papillomavirus (HPV), and persistent HPV infection in particular,
can cause cervical dysplasia (also known as cervical intraepithelial neoplasia [CIN]),
which may
subsequently lead to cancer [1 ]. Although the majority of women with HPV infection will never develop CIN or cancer,
a relatively large
number of women is at risk of developing a CIN.
Nearly all developed countries have implemented cervical cancer screening programs
to reduce the incidence of cervical cancer [2 ]
[3 ]
[4 ]
[5 ]
[6 ]
[7 ]
[8 ]
[9 ]
[10 ]. Due to the introduction of a program in the 1970s, today cervical cancer is a
rare disease in Germany with around 4300 patients diagnosed with cervical cancer every
year [11 ]. The incidence of cervical
intraepithelial neoplasia (CIN 1–3) is 50–100 times higher and often requires colposcopic
evaluation [12 ]
[13 ].
Especially since the beginning of 2020 with the introduction of the new cervical cancer
screening algorithm in Germany, colposcopy plays a major role. Even patients with
one-time low-grade
cytologic anomalies are referred for colposcopy if they have had a high-risk HPV infection.
Patients with persistent HPV infection over one year are also referred [14 ]. In cases of inadequate colposcopy (e.g., scars, bleeding or inflammation), suspected
intracervical CIN2+ or a history of treatment for
cervical dysplasia, the German national guidelines recommend excisional treatment
of the cervix (diagnostic LEEP) [15 ]. The guidelines
also state that normal endocervical curettage (in patients with T3 transformation
zone) does not reliably rule out the presence of a CIN3+, especially in older patients,
and that therefore
diagnostic LEEP should be considered in these cases. We were of the opinion that the
recommendations still leave a lot of questions unanswered, such as: when should cervical
CIN2+ be
suspected? When exactly should diagnostic LEEP be performed in elderly patients with
T3 transformation zone and normal endocervical curettage? In practice, discrepancies
between high-grade
cytological abnormalities and colposcopy-directed biopsies or endocervical curettage
are usually taken into account. The indication for a diagnostic LEEP should be highly
restricted as it
entails an invasive surgical procedure for the patient under general anesthesia which
carries specific surgical risks. Therefore, in order to keep morbidity rates low,
there is a high need for
additional markers which can help clinicians make the right indication for diagnostic
LEEPs. Our retrospectively designed study of patients who were seen in our standardized,
highly frequented
and quality-controlled dysplasia (DKG-certified) unit aimed to find predictors for
the presence of cervical dysplasia in diagnostic loop excisions of the cervix.
Materials/Methods
Study population
The study was designed as a retrospective analysis of all patients who underwent diagnostic
loop excision between 2015 and 2020 at the university hospital in Aachen. Diagnostic
loop
excisions were defined as loop excisions in patients who underwent colposcopy with
biopsy or endocervical curettage (in cases with T3 transformation zone) prior to the
LEEP but without proof
of high-grade intraepithelial lesions. In order to include all patients who underwent
LEEP without prior proof of HSIL, we retrospectively investigated 849 consecutive
patients who underwent
LEEP at University Hospital Aachen between 2015 and 2020. During the study period,
only LEEP was performed, and cold-knife conization was not carried out.
Data collection
In our department, colposcopies are performed in standardized conditions using a Leisegang
3MCV colposcope. The general assessment is carried out using the 2011 International
Federation for
Cervical Pathology and Colposcopy (IFCPC) Terminology for the cervix, with transformation
zone types classified accordingly as 1, 2, or 3. A conventional Pap smear of the cervix,
a test for
human papillomavirus (PCR for HPV DNA) and the application of 5% acetic acid to the
cervix represent the standard of care in our unit, and this procedure is carried out
for every woman
referred with abnormal cytology. During the whole period of investigation, the Seegene
Anyplex II HPV 28 detection kit was used. It simultaneously detects 19 high-risk and
intermediate-risk
HPV genotypes and 9 low-risk types. The classification of HPV viruses into different
categories was in accordance with the IACR (international Association of Cancer Registries)
guidelines.
The detection of multiple high-risk HPV viruses was defined as multiple high-risk
HPV infection. The colposcopic findings were classified in accordance with the IFCPC
terminology as
“normal,” “minor,” “major,” or “suspicious for invasion/cancer.” For cases with T3
transformation zone where no parts of the transformation zone could be visualized
even with splaying of the
cervix, we used the term “non-satisfactory” colposcopy. Normal findings included,
for example, metaplasia, viral warts, and polyps. Minor findings were defined as delicate
punctation, thin
acetowhite epithelium, and irregular geographic borders. Typical major lesions are
represented by sharp borders, an inner border, ridge sign, dense acetowhite epithelium,
a coarse mosaic
pattern, and coarse punctation. Atypical vessels, fragile vessels, irregular surface,
exophytic lesions, necrosis, and ulceration are suspicious for invasion [16 ]. A colposcopy-directed biopsy was taken from the most suspicious part of the lesion.
In some patients with multifocal lesions, more
than one biopsy was taken. In cases with T3 transformation zone, endocervical curettage
was performed. In addition, known risk factors for the presence of cervical intraepithelial
dysplasia
were collected for every patient (see [Table 1 ]), e.g., smoking, immunosuppression, history of LEEP, and patients’ age. All operations
were
performed under colposcopic view of the cervix and were carried out by experienced,
highly qualified, AG-CPC certified staff at the DKG-certified colposcopy unit of University
Hospital
Aachen. Decisions regarding surgical treatment were based on the German S3 guideline
for the prevention of cervical cancer. All included patients had a suspicious cytology
on referral (LSIL;
HSIL; ASC-H, AGC or carcinoma) and normal findings in the colposcopy-directed biopsies
or endocervical curettage. This study aimed to find reliable predictors for the presence
of cervical
dysplasia in diagnostic loop excisions of the cervix.
Table 1
Patient characteristics.
CIN2+
Non-CIN2+
P value
Patients
57 (45.6%)
• CIN 2: 18 (14.4%)
• CIN 3: 38 (30.4%)
• CA: 1 (0.08%)
68 (54.4%)
Age (years)
42.88 ± 10.74
51.40 ± 12.49
< 0.001
History of loop excision
5 (8.77%)
7 (10.29%)
1
Smoker
0.296
1. Yes
24 (42.1%)
19 (27.94%)
2. No
21 (36.84%)
30 (44.14%)
3. History of smoking
9 (15.79%)
14 (20.58%)
4. No information
3 (5.26%)
5 (2.16%)
Immunosuppression
2 (3.5%)
2 (2.9%)
1
Cytology on referral
0.016
• IIID (LSIL)
17 (29.82%)
35 (51.47%)
• IIIg (AGC)
1 (1.75%)
2 (2.94%)
• IIIp (ASC-H)
3 (5.26%)
5 (7.35%)
• IVap (HSIL)
35 (61.40%)
24 (35.29%)
• V (carcinoma)
0
1 (1.47%)
Control cytology
< 0.001
• I/II
9 (15.79%)
26 (38.24%)
• IID1 (LSIL)
8 (14.04%)
19 (27.94%)
• IIID2 (HSIL)
16 (28.07%)
16 (23.53%)
• IIIg (AGC)
2 (3.51%)
0
• IIIp (ASC-H)
2 (3.51%)
3 (4.41%)
• IIIx (AGC favor neoplasia)
3 (5.26%)
0
• IVap (HSIL)
14 (24.56%)
4 (5.88%)
• V (carcinoma)
1 (1.75%)
0
• missing
2 (3.51%)
0
Result of the colposcopy examination
< 0.001
• major change
40 (70.18%)
18 (26.47%)
• minor change
5 (8.77%)
17 (25%)
• unsatisfactory
11 (19.30%)
33 (48.53%)
• missing
1 (1.75%)
0
Transformation zone
0.002
• T1
11 (19.30%)
13 (19.12%)
• T2
20 (35.09%)
7 (10.29%)
• T3
25 (43.86%)
47 (69.12%)
• missing
1 (1.75%)
1 (1.47%)
HPV infection
0.141
• high-risk (non-specified)
4 (7.02%)
2 (2.94%)
• high-risk non-16/18
25 (43.86%)
31 (45.58%)
• high-risk 16/18
22 (38.60%)
19 (27.94%)
• intermediate-risk
2 (3.51%)
4 (5.88%)
• low-risk
0
2 (2.94%)
• negative
2 (3.51%)
10 (14.71%)
• missing
2 (3.51%)
0
Multiple high-risk HPV infection
20 (35.09%)
9 (13.24%)
0.003
Ethical approval
The study was approved by the Ethics Committee of the RWTH Aachen University Faculty
of Medicine in May 2021 (EK 182/21). All procedures performed in this study involving
human participants
were in accordance with the ethical standards of the institutional and/or national
research committee and the 1964 Helsinki Declaration and its later amendments.
Statistical analysis
Continuous variables are expressed as mean values ± standard deviation (SD). Categorical
data are presented as absolute frequencies and percentages. Differences for each variable
between
CIN status (CIN2+ or non-CIN2+) groups were summarized by descriptive statistics and
investigated by Mann-Whitney U-test or Fisher’s exact test.
For the primary analysis, the CIN status was regarded as the primary endpoint. To
identify the most meaningful predictor variables for the CIN status and classify patients
according to
these predictor variables, multivariate logistic regression was performed and a machine-learning
method was used.
Logistic regression was conducted to investigate the influence of known risk factors
(e.g., age, smoking, immunosuppression, history of LEEP) and typical results of the
intensified
gynecological work-up (e.g., colposcopy, transformation zone, HPV testing and cytological
results) on the primary endpoint (CIN2+ or non-CIN2+). Model selection was performed
using the
Akaike information criterion (AIC) [17 ]. The AIC is defined as 2k−2ln(L), where k is the number of predictor variables and
L is the
maximum value of the likelihood function of the logistic regression model. The better
the model fits the data, the higher the value of the likelihood function and thus
the lower the AIC.
Models with a lower AIC are better. The number of predictor variables is positive,
and therefore the 2k-term is often referred to as a “penalty term”, as the addition
of extra variables is
“penalized”, discouraging overfitting. The best-fitting model with the lowest AIC
is described in the results section. The area under the curve (AUC) of the receiver
operating
characteristics curve (ROC) was derived to assess the predictive performance of the
final logistic regression. The final logistic regression included 113 patients; 12
subjects were not
included due to incomplete information (data).
In addition, recursive partitioning was used to generate a decision tree. Recursive
partitioning first finds the best split (CIN2+/non-CIN2+) for all possible covariates
and the splitting
criteria for that covariate and then recursively applies the same procedure for both
new subgroups [18 ]. The goodness of each split is
defined by the “purity” of the new subgroups, i.e., the relative frequency of correct
classifications. The available algorithms differ by how the covariate/split point
is estimated and when
the algorithm terminates. In our analysis, we chose the CART algorithm with a control
parameter maximum depth of any node in the final tree of four, after investigating
different complexity
parameters (cp) and cross-validation results.
We used recursive partitioning as the white-box algorithm in the field of explainable
artificial intelligence (XAI) for several reasons. An additional analysis with a completely
different
approach is useful to investigate the robustness of results, in particular whether
the same variables are identified as meaningful predictors. Furthermore, recursive
partitioning is able to
model non-linear relationships between predictors and outcome (CIN status) well, while
logistic regression, by its very nature, is a linear model. Moreover, in contrast
to black-box
algorithms in the field of machine learning/artificial intelligence such as neural
networks, recursive partitioning yields high explainability, an important feature
in AI (artificial
intelligence). The final decision tree was based on the same dataset that was utilized
for the final logistic regression model. To prevent over-parametrization and allow
better application
of the results to new patients (“unseen data”), the decision trees underwent a procedure
called “pruning”. The optimal pruning parameter was identified by thorough cross validation.
To
assess performance metrics for unseen data, the “leave-one-out” method was used: for
each subject, the algorithm was trained on the remaining 112 subjects and the resulting
algorithm
utilized to classify the subject who was left out during the training. This procedure
was repeated for all 113 subjects. The relative frequency of correct classifications
is referred to as
the accuracy.
A subgroup analysis investigated the distribution of multiple high-risk HPV infections
and the result of the colposcopy examination in different HPV-infection subgroups,
utilizing the
relative row frequencies of the respective contingency tables.
All tests were two-sided and assessed at the 5% significance level. Because of the
exploratory nature of the study, the significance level was not adjusted to account
for multiplicity. All
statistical analyses were conducted using the statistical software R [19 ].
Results
A total of 849 patients with an indication for loop excision of the cervix were assessed
for eligibility. A total of 125 patients without prior proof of CIN2+ were included
in the study.
Inclusion criterium was loop excision without prior proof of high-grade intraepithelial
lesion. [Table 1 ] shows the characteristics of all
included patients and [Fig. 1 ] shows the indications for diagnostic LEEP. Patients with CIN2+ based on diagnostic
LEEP were significantly younger
than patients classified as non-CIN2+. Significant differences between the two groups
were also identified for control cytology, the result of the colposcopy examination,
the transformation
zone, and the presence of multiple high-risk HPV infection, respectively.
Fig. 1
Patient flow chart showing the indications for diagnostic LEEP.
Results of the multivariate analysis
For the multivariate analysis, patients who were PAP IIIg and IIIp on referral were
combined into one category as were PAP IIIg, IIIp and IIIx in control cytology.
Model selection based on the AIC revealed that smoking, a history of LEEP, immunosuppression
and high-risk HPV 16/18 infection did not contribute sufficiently to the model fit,
thus the
applied algorithm determined the final model.
Based on the final multivariate logistic regression model, multiple high-risk HPV
infection (p = 0.001), T2 transformation zone (p = 0.003), major lesion change (p = 0.015)
as the result of
the colposcopy examination and control cytology, in particular the change from category
I/II to IIIg/p/x (p = 0.03), were all found to be statistically significant for CIN
status based on
the result of the diagnostic LEEP.
Furthermore, the final multivariate logistic regression model could be utilized to
predict CIN status. Logistic regression models the probability of CIN2+ in a linear
fashion, expressing
the relationship between predictors and outcome in a closed (i.e., direct and explicit)
form. For our final model, the following formula indicates the probability (p) of
CIN2+:
p=exp(Z)/exp(1+Z), with
Z = −2.309−0.049 × age − 0.716 × cytology on referral Pap III gp + 1.062 × cytology
on referral Pap IV ap + 2.206 × major lesion change + 0.646 × non-satisfactory colposcopy
− 0.175 ×
control cytology PAP III D1 + 0.636 × control cytology PAP III D2 + 2.412 control
cytology PAP III g/p/x + 1.662 × control cytology PAP IV ap + 2.755 T2 transformation
zone + 1.421 T3
transformation zone + 2.195 multiple high-risk HPV infection (see [Table 2 ])
Cytology on referral, control cytology, colposcopy and transformation zone are dummy-coded;
thus, if neither of the mentioned categories is true, e.g., control cytology is I/II,
a 0 has to
be inserted for all mentioned categories. For example, a 60-year-old patient with
cytology on referral Pap IIIg/p, a major lesion change, control cytology PAP III g/p/x,
a T2 transformation
zone and no high-risk multiple HPV infection has an estimated probability of CIN2+
of 80%. The same patient with high-risk multiple HPV infection has an estimated probability
of CIN2+ of
97.29%.
The overall predictive performance of this multivariate logistic regression model
was analyzed using a ROC curve. The AUC value was 88.35% (see [Fig. 2 ]).
Fig. 2
Receiver operating characteristics curve of the final multivariate logistic regression.
Table 2
Final multivariate logistic regression model.
Coefficient
p-value
Intercept
−2.309
0.189
Age
−0.049
0.066
Cytology on referral Pap IIIg/p
−0.716
0.515
Cytology on referral Pap IVap V
1.062
0.065
Major lesion change
2.206
0.012
Non-satisfactory colposcopy
0.646
0.477
Control cytology PAP IIID1
−0.175
0.820
Control cytology PAP IIID2
0.636
0.375
Control cytology PAP III g/p/x
2.412
0.030
control cytology PAP IVap
1.662
0.070
T2 transformation zone
2.755
0.003
T3 transformation zone
1.421
0.071
Multiple high-risk HPV infection
2.195
0.001
Recursive Partitioning
The “leave-one-out” algorithm was found to have an accuracy of 75%, meaning that three
out of four future patients will be classified correctly with regard to the development
of CIN2+.
The final decision tree based on the same dataset that was utilized in the final logistic
regression model is shown in [Fig. 3 ]. The branches
of the tree should be followed to classify whether a patient is CIN2+ or not. Thus,
at the start of the tree, if the colposcopy examination shows a minor lesion change,
the left branch is
followed as the colposcopy findings are classified as either “minor change” or “non-satisfactory”
(i.e., the “yes” path is chosen). In the underlying dataset, 53% of subjects had a
minor
lesion change or non-satisfactory colposcopy, and 75% of this subgroup were non-CIN2+
while 25% were CIN2+. This information is provided in the first knot on the left.
The final leaves of
the tree are determined by following the tree’s path all the way to the bottom, using
“yes” answers on the left or “no” on the right. The purity of the final leaves is
determined
analogously, e.g., 93% of subjects with a major lesion change and multiple HPV infection
were CIN2+.
Fig. 3
Final decision tree of recursive partitioning classifies patients into two CIN-status
classes: CIN2+ and non-CIN2+.
Subgroup analysis for HPV
The subgroup analysis for HPV status showed that multiple high-risk HPV infection
was considerably more common in the subgroup of patients with high-risk HPV type 16
or 18 than in the other
HPV risk categories ([Table 3 ]).
Table 3
Relative frequencies for multiple HPV infection according to HPV status. The subgroup
Lower Risk combines the HPV subgroups negative, LR and
IR.
Multiple HPV infection
No multiple HPV infection
Lower Risk
0.000
1.000
HR non-16/18
0.196
0.804
HR 16/18
0.439
0.561
Colposcopy findings of a major change lesion was more common in patients with high-risk
HPV ([Table 4 ]).
Table 4
Relative frequencies of colposcopy results by HPV status. The subgroup Lower Risk
combines the HPV subgroups negative, LR and IR.
Minor lesion change
Major lesion change
Non-satisfactory colposcopy
Lower Risk
0.158
0.211
0.632
HR non-16/18
0.196
0.482
0.321
HR 16/18
0.171
0.537
0.293
Discussion
In the literature, the reported accuracy of colposcopy-directed biopsy ranges from
60–95% for HSIL [20 ]
[21 ]. The histological differences between colposcopy-directed biopsy and loop excision
of the transformation zone (LETZ) have been a cause
for concern for a long time [22 ]
[23 ]
[24 ]. The German S3 guidelines list specific indications for carrying out diagnostic
loop excisions. In our opinion, the patient cohort with indications for carrying out
diagnostic LEEP in accordance with the S3 guidelines is quite large, especially as
it includes elderly patients with T3 transformation zone and negative endocervical
curettage, and in most
colposcopy clinics, the cohort has increased considerably since the start of the new
cervical cancer screening in Germany. But it also includes patients with suspected
endocervical CIN2+ and
discrepant cytology, colposcopy and histology findings. This study aimed to narrow
down these groups and find evidence-based influencing factors for the presence of
CIN2+ in diagnostic LEEPs
[25 ]
[26 ].
The majority of studies evaluate the concordance rates for preoperative colposcopy-directed
biopsies and cone histology. Duesing et al. included 36 patients in their analysis
of 166 patients
without preoperative detection of CIN2+. In the accuracy analysis of colposcopy-directed
biopsies by Stuebs et al., 106 of 642 patient had normal/LSIL findings in the colposcopy-directed
biopsies. Thus, “diagnostic LEEP” rates were 21% and 16%, respectively [25 ]
[27 ].
This is in line with our diagnostic LEEP rate of 14.7%. Unfortunately, no further
attention has been given so far to the characteristics of the patient group with negative
biopsies prior to
LEEP surgery, although this patient collective seems to exist in all colposcopy centres.
In the systematic meta-analysis of Underwood et al., the sensitivity of colposcopy-directed
biopsies
for detecting CIN2+ was 80.1%. Thus, CIN2+ was found in excisional biopsies (LEEP)
in 20% of cases without prior detection via colposcopy-directed biopsies. A multivariate
analysis of
potential factors affecting the quality of colposcopy-directed biopsies was not possible
because of insufficient data in most of the included studies [20 ].
We present here a study based on our retrospective data showing that HPV high-risk
multiple infection is a key indicator for the presence of HSIL in diagnostic loop
excisions. In the final
multivariate logistic regression model, the presence of a T2 transformation zone rather
than T3 or T1, colposcopy findings of a major lesion change and the presence of multiple
high-risk HPV
infections had a major influence on the presence of HSIL in diagnostic loop excisions.
Interestingly, the machine-learning technique (recursive partitioning) identified
similar variables as
important for CIN-status classification; thus, the results can be regarded as robust.
In cases with negative colposcopy findings (only minor changes or unsatisfactory colposcopy)
in patients
aged ≥ 44, the estimated probability that diagnostic loop excision will be negative
(without a histological finding of CIN2+) was 84%. On the other hand, in cases with
a major colposcopic
lesion change and the presence of a multiple high-risk HPV infection, the risk of
HSIL is 93% (See [Fig. 3 ]). According to the ASCCP guidelines,
the NHSCS, and the German guidelines, excisional procedures are not recommended in
patients with CIN 1 or normal biopsies [15 ]
[28 ]
[29 ]. Exceptions can be made in cases with suspected endocervical dysplasia or
unsatisfactory colposcopy in combination with abnormal cytology findings. Accordingly,
57.6% of our patient collective had a non-visible transformation zone (T3) and 35.2%
an unsatisfactory
colposcopy, with a non-visible transformation zone, even upon splaying of the cervix.
In practice, colposcopists often face a diagnostic challenge when dealing with patients
with a T3
transformation zone. Nevertheless, diagnostic discrepancies also exist in patients
with T2 or T1 transformation zone. Of course, in these cases it is rarer, as colposcopy-directed
biopsies can
be taken with greater accuracy [27 ]. Interestingly, the presence of a T2 transformation zone rather than T3, and the
presence of a major
colposcopic lesion change were the main predictors of the presence of CIN2+ in our
study. Thus, looking at our data, the presence of diagnostic discrepancies in patients
with T2 or T1
transformation zones is a stronger predictor of CIN2+ than in patients with a T3 transformation
zone. This can be explained by the higher diagnostic value of colposcopy in patients
with T1 or
T2 transformation zone. In cases of major colposcopic lesion changes, our data suggests
that when this is combined with the presence of other risk factors (e.g., multiple
high-risk HPV
infection), diagnostic loop excision should be considered even when biopsies are negative.
Treating major colposcopy-detected lesion changes without additional risk factors
would lead to an
overtreatment of patients as shown in the TOMBOLA study [30 ].
The overwhelming majority of sexually active women and men have been infected with
HPV at least once in their lifetime [31 ]. A woman
might have been infected with two HPV types by one partner and become infected with
a third HPV type later on. One infection can persist even after several others have
cleared [32 ]. In our study, we found that the presence of multiple high-risk HPV infection is
a key predictor of the presence of CIN2+ in diagnostic
loop excisions. We performed a sub-analysis for HPV status which showed that infection
with high-risk HPV type 16 or 18 (high-risk HPV 16/18 group) is associated with multiple
high-risk HPV
infection and major colposcopic lesions ([Table 3 ] and [Table 4 ]). Multiple high-risk HPV infection
and major colposcopic lesion changes are far more common in the high-risk HPV 16/18
subgroup. Multivariate analysis revealed that, given information on multiple high-risk
HPV infection and
major change lesion, the HPV risk status does not have an incremental value and thus
does not contribute notably to the model fit. Due to this confounding, the HPV risk
status correctly does
not appear in the final model, although it is known to have a high carcinogenic potential
[32 ]. Of course, the colposcopist needs a
modern HPV DNA detection kit which can identify the different HPV genotypes and classify
them in different categories in accordance with the IACR (International Association
of Cancer
Registries) guidelines. When using tools that only give binary information (HPV 16/18
or other), information which would be of value to the gynecologist (multiple high-risk
HPV infection) is
unavailable.
Our study has several strengths and limitations that need to be addressed. It is a
retrospective study with a limited patient cohort. It will be very interesting to
see whether the identified
influencing variables will be confirmed in a prospective multicenter study.
The overall predictive performance of this multivariate logistic regression model
is very good, as demonstrated by the high AUC value of 88.35% of the ROC curve. By
using machine-learning
algorithms on our data, we were able to show that the accuracy for unseen data is
75%, which is relatively high despite the limited patient cohort.
Conclusion
Our data showed that high-grade cytological abnormalities (PAP IV – HSIL), neither
upon referral nor in the control cytology have a major influence. Clinicians should
rather focus on the
results of the colposcopy examination (T2, major change) and of HPV testing (multiple
high-risk HPV infection) when considering diagnostic excisional procedures of the
cervix.
