Abstract:
Chronic osteomyelitis, predominantly affecting the tibia bone stems from an inadequate
treatment of post-fracture, with high prevalence in developing countries. This study endeavors to
classify chronic tibia shaft osteomyelitis as type 1 (medullary), type 2 (cortical) type 3
(superficial) and type 4 (diffused) based on the spread of infection. In-vitro samples of normal
and chronic tibia shaft osteomyelitis were collected in collaboration with orthopedists from
Jimma University Medical Center and Woliata Soddo University Referral Hospital. With a
customized analog front end circuit, the Bio impedance measurement was carried out using a
four-electrode system with an inter electrode distance of 3cm. The electrode were connected to
an AD5933 impedance evaluation board and the impedance was measured on the resized bones
of length, 4cm- 6cm. A complete dataset consisted of 11 discrete frequencies within 2 kHz to 99
kHz with an acquisition period lasting 7 minutes. We validated the accuracy of the prototype
using a phantom circuit containing the resistor and the capacitor that resembled the impedance
characteristics of the five different classes of tibia shaft tissues, namely the normal, chronic tibia
shaft tissue types (type 1, type 2, type 3 and type 4). We obtained an average accuracy of 99.9%
for the validation. Data analysis based on Nyquist plot, Bode plot, and Cole-Cole plots showed
discernible patterns for infected tibia shaft tissue. Higher stability, permittivity and ability to
store charges have been observed for normal tibia shaft; whereas these properties were
progressively decreasing for the osteomyelitis type 1,2,3,4. In general the impedance decreases
with increase in frequency, as has been observed in our cases too. But the osteomyelitis bones
exhibited lesser impedance drop compared to the normal bones with increasing frequency. For
example the impedances measured at 2 kHz and 99 kHz for different types of bones are as
follows: Normal: 244.4624 Ω and 62.5 Ω, Chronic shaft tissue osteomyelitis type 1: 248.1772 Ω
and 110.4722 Ω, type 2: 284.1804 Ω and 128.3072, type 3: 595.5715 Ω and 470.3853 Ω, type 4:
2134.332 Ω and 774.1446 Ω. To make our prediction robust and automated, seven different
classification models were developed using different machine learning algorithms. Among these,
the Extra Tree algorithm based model performed well with the highest accuracy of 98%.
Subsequent testing with biopsy samples also yielded an accuracy of 98%
